Gonçalo Esteves / building capable systems Field Notes →

The Vault / Companies

Company

Anthropic

the AI lab the panel talks about more than any other and had no page for — 72 of 88 indexed Moonshots with Peter Diamandis episodes, 1,324 mentions, and on Moonshots 288 — Three Lab Warnings in Five Days, Researcher Flags “Gambling with Our Lives,” and Labs Race alone 35 mentions

What it is (one line): the AI lab the panel talks about more than any other and had no page for — 72 of 88 indexed Moonshots with Peter Diamandis episodes, 1,324 mentions, and on Moonshots #288 alone 35 mentions across seven separate segments that are never framed as being about a company.

Updated 2026-11-16. Refreshed weekly (see How this page stays current at the bottom). New on this date: an 11 to 16 Nov This week block. The IPO terms did not change. No third S-1/A, no effectiveness notice and no final prospectus were on EDGAR on 16 Nov. What was filed is five free-writing prospectuses (10 to 12 Nov). Anthropic's own IPO page (ipo.anthropic.com) says the price is "expected to be set on November 18th" and trading is "expected to begin" on 19 Nov. The demand side is reported, not filed: the book was about six times covered by 12 Nov (Reuters and Bloomberg, unnamed sources), and the FT says bankers are weighing a price above the $70–78 range. Correction to the last two refreshes: the government's answer in the D.C. Circuit is not due 13 Nov. A 3 Nov clerk's order vacated the 23 Oct order and gave the government 15 days, so the answer is due about 18 Nov, the reported pricing day. A think-tank amicus opposing rehearing was filed 9 to 10 Nov. Products: Claude Meetings (beta), Claude for Manufacturing and self-hosted Managed Agents, all on 12 Nov; the 2026 usage policy took effect the same day. Research: a safety-evaluations guide (13 Nov) and a shorter vulnerability-disclosure window (16 Nov: 90 days down to 30 or 45, from 1 Jan 2027). New and open: Claude for Open Source, six months of Claude Ultra at no cost for maintainers and contributors who meet one of five printed thresholds, with no deadline printed. The Residency still closes 22 Nov.

Previous update, 2026-11-10: a 5 to 10 Nov block. The IPO has a price. A second S-1/A was filed on 10 Nov (EDGAR), and the same day Anthropic's roadshow release set the range at $70.00 to $78.00 a share for 825,000,000 Class A shares: 801,351,354 from the company and 23,648,646 from selling holders, plus a 123,750,000-share option for the underwriters. Pricing is reported for 18 Nov, with trading on 19 Nov (CNBC and Bloomberg, unnamed sources). The reported valuation depends on the basis used: up to ~$1.4T fully diluted (CNBC) against up to $1.02T (Reuters). Claude Fable 5.2 and Mythos 5.2 were released on 9 Nov at $8 / $40 per million tokens, down from $10 / $50. Also this week: METR's outside assessment (5 Nov, ~1.9× faster capability gains), a $3.4B gas-generator purchase (FT, 8 Nov, not announced by the company) and five research posts. Courts: the government's answer to the rehearing petition is still due 13 Nov (corrected 16 Nov: a 3 Nov order moved it to ~18 Nov). New and open: the Anthropic Residency, a paid AI-safety research placement of six weeks to six months for professors, founders and staff engineers, with no ML background required. Applications close 22 Nov 2026, 11:59 PM PT.

Previous update, 2026-11-04: a 30 Oct to 4 Nov block. The first amendment is filed: an S-1/A on 2 Nov (EDGAR; price range and share count still blank; secondary readers of it say it adds Q3 — revenue $17.0B — and finally quantifies the compute commitments the S-1 left blank at **$895B**). SpaceX's Q3 10-Q answers this page's watch item: an unnamed "Customer B" paid 34.5% of SpaceX's Q3 revenue (~$5.0B), and Bloomberg reports that customer is Anthropic — reportedly now SpaceX's largest customer, ahead of the US government. The D.C. Circuit gave the government to 13 Nov to answer the rehearing petition (order of 23 Oct, found this run; vacated by a 3 Nov order, see the 16 Nov update). Two research posts (provable inference, 2 Nov; vault particles, 4 Nov); no newsroom item since 28 Oct. New and open: nothing new. The Fellows deadline (30 Oct) has passed and its page now 404s; the guest-investigator call still prints no closing date and was still open today.

Why it is in this archive

Gate: cited in 72 episodes — the single largest hole in the graph (the research archive, promotion candidates). Promoted on 2026-09-23 off the EP288 decomposition, where it is simultaneously the safety story, the governance story and the market story of the episode.

Facts

Founded 2021
Revenue $6.5B quarterly run rate ($26B annualised) — per an NPR deep-dive read on Moonshots #288 1:54:48 ⚠️ not independently checked · superseded by reporting, 2026-10-15: Q3 2026 revenue $16.8B (Q2 ~$11.6B, Q1 $4.7B), told to investors on 14 Oct and reported by the FT, Reuters (two attendees) and Bloomberg; ⚠️ not published by Anthropic, not in a filing · now a filed figure, read secondhand (2026-11-04): the 2 Nov S-1/A reportedly states Q3 2026 revenue $16,956M and nine-month revenue $33,242M — quoted from the filing by others; the amendment itself has not been opened by this page
Market position "42% of the AI coding market", a "$35 billion cloud deal", NVIDIA-backed infrastructure — same NPR source ⚠️ · the $35B figure matches the Lambda cloud deal reported by the WSJ and Reuters on 31 Aug / 1 Sep 2026 from unnamed sources (~350 MW, Nueces County, Texas); neither company has announced it
IPO ✅ Public S-1 filed 2026-10-26 (EDGAR, CIK 0002130357, File No. 333-299633; anthropic.com): Class A common stock, Nasdaq Global Select Market, symbol "ANT"; "The number of shares … and the price range … have not yet been determined"; first S-1/A filed 2026-11-02 (accession 0001628280-26-068676) — price range and share count still blank in it (per every filing-based secondary read; index checked at EDGAR); second S-1/A filed 2026-11-10 (accession 0001628280-26-072484, accepted 06:29, 65 documents, adds an EX-1.1 underwriting agreement and a fee exhibit) and the price range is set the same day: "The initial public offering price is expected to be between $70.00 and $78.00 per share", 825,000,000 shares (801,351,354 primary, 23,648,646 secondary), 123,750,000-share option (anthropic.com, primary; main S-1/A document not opened). Five free-writing prospectuses (FWP) filed 10 to 12 Nov under the same file number (EDGAR; filenames point to the IPO website, UK materials, an Australian wrapper, an EEA prospectus and a Japan amendment; the documents were not opened). As of 16 Nov there is no third S-1/A, no 8-A, no effectiveness notice and no 424B4. ipo.anthropic.com (company): price "expected to be set on November 18th", trading expected 19 Nov, "The registration statement has not yet become effective"; Q3 2026 revenue $17.0B, net loss $23.3B including a $21.9B non-cash charge. Valuation at the top of the range is reported as up to ~$1.4T fully diluted (CNBC) or up to $1.02T (Reuters); the bases differ, and neither is a company figure. The financial figures in coverage are secondary readings of the filing (see This week) · "a mega IPO approaching at 2 trillion or more" ⚠️ on-tape characterisation, not a filing · earlier: confidential draft S-1 submitted 2026-06-01 (anthropic.com); (history) no public S-1 on EDGAR as of 2026-10-23 (full-text search, "Anthropic PBC", 17 to 23 Oct: 4 fund filings, none by Anthropic) · an updated draft was reported by the WSJ on 20 Oct (H1 2026 revenue $16.3B, commitments >$600B; ⚠️ leaked, not filed; see This week) · a prospectus was reported out ~2026-09-29 (see Two weeks back, 2026-10-06 row) · investor day 14 Oct 2026 (reported); marketing "as early as the week of November 9", trading before 26 Nov (Bloomberg via secondary, "plans remain fluid")
Comparison drawn on air "it took Google 8 years to reach that revenue. Anthropic did it in five." (Moonshots #288 1:55:11)
Models named in one episode Claude Fable 5.1 (public), Mythos 5.1 (restricted access), "Opus". Keep these apart — the panel does not.

The economic impact report — the numbers that actually travel

Anthropic published its own economic impact report ahead of the IPO (Moonshots #288 1:55:22). Three scenarios against a stated $30 trillion total addressable market. The extreme scenario, as read on air:

variable extreme scenario
share of today's cognitive work done by AI by 2030 ~half
GDP growth 15% a year
labour share of income 60% → 45%
cognitive workers unemployed ~1 in 5

"Even the moderate scenario shows meaningfully faster GDP growth and substantial white collar displacement." (Peter, 1:56:14) Anton Korinek is named as an author (2:04:20).

Both economists on the panel say the report is wrong in opposite directions. Alexander Wissner-Gross: "I think anthropic is lowballing estimates… I expect 2x or 3x year-over-year growth and not just 15." (1:58:21) Emad Mostaque: "we're screwed with these numbers and they have not done their model properly… there's a complete collapse in aggregate demand … they have aggregate wages staying constant which just mathematically is impossible." (2:02:14) ⚠️ The 15% figure is a company's own pre-IPO scenario, published by an interested party. It is the most quoted number of the episode. Do not repeat it as a forecast.

✅ Checked against the paper, 2026-09-29. The report exists: Economic Scenarios for Transformative AI, Korinek, Jones, Sacher, Cotter and McCrory, The Anthropic Institute Working Paper No. 2026-02, September 2026 (PDF, scenario explorer). All four numbers in the table above are in its abstract, verbatim: "AI performing almost half of today's cognitive work by 2030. GDP growth then rises to 15 percent per year, the labor share of income falls from 60 to 45 percent, and nearly one in five cognitive workers is unemployed." Three corrections to the on-air version:

  • The paper's scenarios are modest, substantial and extreme. There is no "moderate" scenario; Peter's "moderate" is the paper's substantial.
  • The "$30 trillion total addressable market" is not in the paper (searched the full text). ⚠️ Unverified; do not attach it to this report.
  • Emad Mostaque's "they have aggregate wages staying constant" is close, but it is an output, not an assumption. The paper: "Total labor income in 2030 in the extreme scenario is almost exactly what it would have been without AI: the labor share falls by a quarter while GDP rises by a third, and 0.45 × 1.32 ≈ 0.60. All of the increase in GDP therefore accrues as capital income." His read of the result is right; his read of the method is not.

This week — 11 to 16 Nov 2026

Checked 2026-11-16 against anthropic.com/news (newest titled item: 10 Nov), anthropic.com/research (newest: 16 Nov), claude.com product updates (newest: 12 Nov), EDGAR (the full filing list for CIK 0002130357), ipo.anthropic.com, the Residency announcement and both postings, the Startups, Economic Futures and Fellows pages, a D.C. Circuit docket tracker (CourtListener itself refused the fetch), and secondary coverage for the book, the pricing talk and the courts. Replaced every week; the 5 to 10 Nov block follows it.

date what source
10–12 Nov IPO filings: five free-writing prospectuses, nothing else. EDGAR lists five FWPs under file no. 333-299633: three accepted on the evening of 10 Nov ET and two on the morning of 12 Nov (all carry a 12 Nov filing date). By their filenames they cover the IPO website, UK materials, an Australian wrapper, an EEA prospectus and a Japan amendment. The documents were not opened. As of 16 Nov there is no third S-1/A, no Form 8-A, no SEC effectiveness notice and no 424B4 final prospectus, so the $70–78 range and 825,000,000 shares stand. EDGAR filing list, primary (index)
~11-12 ipo.anthropic.com goes up (announced on X on 12 Nov, per one secondary roundup). The page says the final price is "expected to be set on November 18th", subject to regulatory clearance, and trading is expected to begin 19 Nov on Nasdaq as ANT. It names Schwab, E*TRADE, Fidelity, Robinhood and SoFi as brokers and Goldman Sachs and Morgan Stanley in the legal notice, and says "The registration statement has not yet become effective." Its Q3 2026 highlights: revenue $17.0B; net loss $23.3B, including a $21.9B non-cash charge; adjusted operating margin 15%; operating margin (3)%. The prospectus, fact sheet and founders' letter PDFs it links were not opened. The page prints no price range. ipo.anthropic.com, primary (company)
11-11 → 11-13 Demand, as reported (none of it filed or announced): the book was "fully covered" when the deal launched (Bloomberg). By the end of 10 Nov it was about 2× (Reuters, one source), "multiple times" by 11 Nov (Bloomberg), and about 6× by the end of 12 Nov (Reuters and Bloomberg, unnamed sources). Bankers are reported to be prioritising existing backers (Bloomberg). The FT (11 Nov) says bankers are weighing a price above the $70–78 range and a valuation of $1.5T or more. Reuters (13 Nov) says the roadshow runs to Wednesday 18 Nov, when the deal is expected to price, with meetings in New York, Boston and San Francisco. Banks model 2028 revenue at ~$270–280B, and some sell-side models put 2026 at ~$52–53B. A JonesTrading block of up to $500M of expected allocations is being shopped at a 3–5% fee (Bloomberg, 10 Nov). CNBC (15 Nov): a nine-month net loss of $104B and "over $900 billion" of commitments, per the latest prospectus as CNBC reads it. ⚠️ All of this is reported. A price above the range would need a new filing or a 424B to count here. Reuters via News & Observer, Bloomberg, 11 Nov, Bloomberg Law on the FT, Bloomberg, 10 Nov, CNBC, 15 Nov, secondary
11-03 (found this run) Courts: correction. The two previous refreshes said the government's answer to the 13 Oct rehearing petition was due 13 Nov. That is wrong. A 3 Nov clerk's order vacated the 23 Oct order and, "on the court's own motion", gave the respondents 15 days to answer the petition for panel rehearing and rehearing en banc. That puts the answer at about 18 Nov, the reported pricing day; 18 Nov is our arithmetic, and the tracker does not print the date. One IPO explainer (11 Nov) gives the same date from CourtListener. 9–10 Nov: an amicus brief by Joel Thayer opposing rehearing en banc, with a motion for leave to file it. The America First Policy Institute announced it as its brief, and Law360 calls it a "Trump-aligned think tank". No government response and no ruling appear as of the tracker's latest entry (10 Nov). AI lawsuit tracker (docket entries, "verified 2026-11-16"), AFPI, Law360, ap7i, secondary; docket itself still not read (CourtListener returned 403)
11-12 Products (claude.com): Claude Meetings (beta, Team and Enterprise, Google Meet only so far): Claude joins as a visible participant and keeps "a transcript, summary, and screen descriptions, not audio". It stays out of calls with outside guests by default. Artifacts made in calls are free "through February 10, 2027". It is not available to organisations with zero data retention, HIPAA-ready configuration or customer-managed keys. This settles the EP311 "Claude meetings" item below: it was "coming soon" on 10 Nov and shipped in beta on 12 Nov. Also: Claude for Manufacturing, a plugin with 17 skills and connectors to Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite (no separate price), and self-hosted Claude Managed Agents (private beta, by request through an account team). Opus 5.5 is now the default model on Pro, Max, Ultra and Team (secondary reads of a 12 Nov guide). claude.com: Claude Meetings, product updates, primary; vanemmerik.ai, secondary
11-12 The 2026 usage policy takes effect (announced 8 Oct; see the 2026-10-15 episode section). Nothing new in the text; the date was set in October. anthropic.com, primary (Oct post)
11-13 Research: Building strong safety evaluations (Engineering). Recommendations drawn from the system cards: "100 items as the floor for a complex, multi-turn eval and 1,000 as the floor for a simpler, single-turn one". The same eval scored "60% on the API vs 94% on claude.ai", and configuration differences moved a Fable 5.1 result by as much as 34 points. Self-reported. No call. anthropic.com/research, primary
11-16 Research: a shorter vulnerability-disclosure window (Frontier Red Team). "from 90 days to 30 days for high- and critical-severity issues, and to 45 days for all other issues", applying to reports made "on or after January 1, 2027". Escalation to an outside coordinator comes after 15 days without a maintainer response (was 30), and full technical details follow 30 days after a patch (was 45). The 7-day window for actively exploited bugs is unchanged. The post points maintainers to OSS Scanner and to Claude for Open Source (next row). anthropic.com/research, primary
11-16 (first seen) Claude for Open Source (see What is open now): "6 months of Claude Ultra on us" for anyone meeting one of five thresholds: 500+ dependent repos, 100+ dependent packages or 200,000+ monthly downloads; a listed committer on a recognised foundation or language project; 100+ merged PRs to repos you don't own in 12 months; a repo with 20+ outside contributors in 12 months; or an OpenSSF criticality score of 0.4 or more. "Don't quite fit? … apply anyway." No launch date, deadline or cap printed. New row in submissions/pipeline.json. claude.com, primary
— Newsroom: no new titled item since 10 Nov. The index shows a bare "November 16, 2026 · Announcements" header with no item attached; nothing to read. anthropic.com/news
— Space: no new primary tie. A 13 Nov wire repost of SpaceX's Colossus 1 deal and the "multiple gigawatts of orbital AI compute" interest is the 6 May announcement (already in Anthropic and space), re-dated by the aggregator. Not new. No JPL, NASA or orbital-compute filing names Anthropic this week. search, bounded
— Corpus: anthropic appears in 92 of 111 indexed Moonshots with Peter Diamandis episodes, 1,756 mentions (index through the 13 Nov episode). It is still the most-mentioned company by count (OpenAI 1,423, SpaceX 1,043, NVIDIA 610). By reach it is still behind Meta (96) and OpenAI (93). the research archive

Programmes re-checked 2026-11-16: Residency: open. Both postings are live and still read "Applications for this round close on November 22, 2026 (11:59 PM PT)". New: Claude for Open Source (row above). Fellows: anthropic.com/fellows still 404. Three "Anthropic Fellows Program" postings are still on page 1 of the Greenhouse board, not opened. Claude for Startups: still "not accepting new applications right now". Economic Futures: still "not currently accepting applications". Claude Corps: not re-opened this run. Search still indexes the CodePath Greenhouse posting with the "January 31, 2027" deadline, so the date stands. AI for Science and guest investigators: not re-checked this run (the help article and the call page were not reached). The 10 Nov readings stand: next AI for Science review 7 Dec; the guest-investigator call has no printed close.

What it adds up to: a quiet week on paper and a loud one in the press. Nothing filed changes the terms, but the reported book (about 6× covered) and the FT's talk of pricing above the range point the other way from last week's "lower than the talk". Two dates now fall on the same day, 18 Nov: the reported pricing and, by the corrected count, the government's answer in the D.C. Circuit.

Previous week — 5 to 10 Nov 2026

Checked 2026-11-10 against anthropic.com/news (newest item: 10 Nov), anthropic.com/research (newest: 9 Nov), EDGAR (the full filing list for CIK 0002130357 and the 10 Nov S-1/A index), alignment.anthropic.com and both Residency postings, every programme page in What is open now*, and secondary coverage for the valuation, the pricing date, the generator deal and the courts. Replaced every week; the 30 Oct to 4 Nov block follows it.*

date what source
11-10 Price range set; the roadshow is open. "Today Anthropic, PBC launched the roadshow for its proposed initial public offering"; 825,000,000 Class A shares, of which 801,351,354 are from Anthropic and 23,648,646 from "certain existing stockholders" (Anthropic gets no proceeds from those); a 30-day option for "up to an additional 123,750,000 shares"; "between $70.00 and $78.00 per share"; Nasdaq Global Select, ANT. The second S-1/A reached EDGAR the same morning: accession 0001628280-26-072484, accepted 10 Nov 06:29, 65 documents, now including an EX-1.1 (underwriting agreement) and a filing-fee exhibit. The main document (14.8 MB) was not opened. Arithmetic on the release's own numbers: $57.8B to $64.4B for the base deal (≈$62.5B to the company at the top). What is reported, not filed or announced: pricing on 18 Nov and trading on 19 Nov (CNBC, unnamed sources; Bloomberg says the same from deal terms it reviewed); **20% retail allocation**; a book "already covered" at the range (CNBC sources); a $26B private placement at the IPO price from cornerstone investors including Google, NVIDIA and AMD (Reuters; up to $90.4B in total with the offering). The valuation depends on the basis: up to $1.405T fully diluted (CNBC) against up to $1.02T (Reuters). Both are well below the $1.5T to $2T talk this page carried in October. Late on 9 Nov a version of the filing showing $70–80 and ~866.7M shares went to banks and was retracted as "inaccurate" (Reuters, emails seen). The $70–80 figures are not the terms. anthropic.com, EDGAR S-1/A index, primary; CNBC, NBC News (Reuters), BusinessLine (Reuters), secondary
11-09 Claude Fable 5.2 and Claude Mythos 5.2, "the same underlying model". Fable ships with cyber and biology safeguards and is available to all users; Mythos is for vetted organisations through the Cyber and Life Sciences Verification Programs (Enterprise customers also get it through Claude Security). Price $8 / $40 per million input/output tokens (from $10 / $50); fast mode at $16 / $80, "up to 2.5 times" faster; output "up to 30% faster" than 5.1. FrontierCode 2.0 at Xhigh effort: 49.1% (Fable 5.1 32.4%, Opus 5.5 45.5%, GPT-6 Astra 42.8%). The page also claims the model "formalized the proof of Fermat's Last Theorem in 30 hours" and "solved two open problems in the physics of black holes". Mythos 5.2 CVE Rediscovery: 89.6% vs 78.3% (Computerworld reading the post). Data retention is 30 days for customers without zero-data-retention terms. All benchmarks are self-reported. One outside data-analysis test (Plotly) gave it 43/43 at the second-highest run cost on its board. Access programme, not a new call. anthropic.com, primary, self-reported; Computerworld, Plotly, secondary
11-09 Electronics design (research): working with Diode Computers, Fable 5.2 designed three electronic prototypes that worked without circuit fixes. PCB Gym: "Fable 5.2 scored 69% on the benchmark, compared with 25% for Fable 5". Self-reported. anthropic.com/research, primary
11-08 $3.4B for gas generators (FT). Anthropic took over 1.7 GW of Caterpillar generator orders held by OnePWR Solutions (San Antonio), with installation engineering included and no company acquired. Signed late September; units expected within about two years; sites undisclosed. The FT calls it the largest known single purchase of power-generation equipment by an AI lab. Piper Sandler's September deal note recorded a $3.4B sale of OnePWR's platform to an unnamed "frontier AI lab". ⚠️ Not announced by Anthropic. Both the FT identification and the amount are reported. Bloomberg on the FT, DCD, Fortune, secondary
11-05 METR's outside assessment (Anthropic invited METR in): "As of October 2026, Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025", with "high uncertainty" (1.3× to 3× would not surprise METR). It attributes ≈80% of the acceleration to AI use, "highly uncertain". Details are on the EP309 section below. METR, primary (third party); AP
11-05/06 Research, three posts: How researchers use Claude (5 Nov: "We analyzed 207,000 research sessions"; data analysis is 33.7% of Claude sessions and 48.1% of Claude Code sessions, against 27.6% of researcher time); The components of agent memory (6 Nov: "dreaming" cut a memory store from "413 memory files across 225 folders" to "nine files in two folders"); Teaching AI to build scientific models (6 Nov, a guest essay by AI for Science grantees). All self-reported. No call in any of them. anthropic.com/research, primary
11-05 / 11-10 Products (claude.com, not the newsroom): Personal Finance in Claude (5 Nov, US beta on Pro, Max and Ultra plans: connect bank and card accounts); 10 Nov, Claude Sheets, Whiteboards, Data and interactive connectors, launched at a company event (Bloomberg frames them as competing with Microsoft and Google, both of them backers). claude.com, primary; Bloomberg Law, secondary
~11-03 The Anthropic Residency opens (see What is open now). A paid, fixed-term AI-safety research placement: "Placements run for six weeks, three months, or six months"; "$350,000 to $500,000/year pro-rated" (US) and "£265,000 to £375,000/year pro-rated" (UK); "Background in machine learning is not required"; "Applications for this round close on November 22, 2026 (11:59 PM PT)"; offers expected January 2027; visas sponsored for placements of three months or longer. Neither page shows a posting date; coverage puts it at 2 to 3 Nov. New row in submissions/pipeline.json. alignment.anthropic.com, US posting, UK posting, primary
— Courts: nothing new. No order on the 13 Oct rehearing petition was found. The government's response is still due 13 Nov 2026 (23 Oct order, per the docket tracker). Corrected 16 Nov: a 3 Nov order vacated that one; the answer is due ~18 Nov (see This week*).* The docket itself has still not been read. search, bounded
— Space: no new primary tie. The SpaceX Customer B line (previous week) is still the big one. Secondary coverage this week repeats the ~$3.1B a month at full delivery and the early-2027 exit window. One podcast argument says SpaceX could use the 90-day clause to reprice (All-In, opinion). No new JPL or NASA contract item, and no orbital-compute filing, names Anthropic. search, bounded
— Corpus: anthropic appears in 90 of 109 indexed Moonshots with Peter Diamandis episodes, 1,660 mentions (index through the 7 Nov episode, which is titled for the IPO). It is still the most-mentioned company by count (OpenAI 1,395, SpaceX 1,003, NVIDIA 536). By reach it is still behind Meta (95) and OpenAI (91). the research archive

Programmes re-checked 2026-11-10: New: the Anthropic Residency (row above; closes 22 Nov). Fellows: anthropic.com/fellows and /careers/fellows still 404. The Greenhouse board still lists three "Anthropic Fellows Program" postings, which were not opened. Guest investigators: the Apply button and Google Form are still linked, with no closing date printed ("We aim to decide within 14 days"). The page cannot show whether the form still accepts entries. AI for Science: the help article is unchanged (16 Mar 2026; first-Monday reviews; "up to $20,000"); next review 7 Dec 2026. Claude for Startups: still closed. Economic Futures: still "not currently accepting applications". Claude Corps: not re-verified this run. The 2027 posting was not on pages 1 to 4 of the 14-page Greenhouse board, and the rest were not read. The 31 Jan 2027 deadline (read 29 Oct) stands until it is checked again.

What it adds up to: the two numbers the market was waiting for have arrived, and both came in lower than the talk: a price range that values the company at about $1.0T to $1.4T depending on how you count, against the $2T said on air in September. In the same 48 hours the company shipped a flagship model at a 20% lower price and was reported to own 1.7 GW of its own generators. Pricing is reported for 18 Nov, and the government's court answer is due 13 Nov (corrected 16 Nov: ~18 Nov).

Two weeks back — 30 Oct to 4 Nov 2026

Checked 2026-11-04 against anthropic.com/news (newest item shown: 28 Oct), anthropic.com/research (newest: 4 Nov), EDGAR (the full filing list for CIK 0002130357 and the S-1/A index), every programme page in What is open now*, a D.C. Circuit docket tracker, and secondary coverage for the S-1/A figures, the SpaceX 10-Q and the Nscale talks. Replaced every week; the 24 to 29 Oct block follows it.*

date what source
11-02 The first amendment: S-1/A filed. Accession 0001628280-26-068676, accepted 2 Nov 09:24, 50 documents (one main document, 14.3 MB, not opened this run; four exhibits — a charter document, an instrument, two auditor consents — and 45 graphics; no underwriting agreement, no new material contracts, no fee table). Price range and share count still blank — consistent across every filing-based secondary read. What those readers say it adds: Q3 results (revenue $16,956M, nine months $33,242M; Q3 operating loss $567M, attributed "primarily" to $5.0B of short-term compute and $2.0B of non-cash donation-matching charges; adjusted operating margin 19% → 15%); the commitment line the S-1 left blank is now ~$895B of compute purchase commitments and leases, **$69.8B due within 12 months**; a new "Sales to Retail Investors" section naming Schwab, Fidelity, Robinhood, Trade-PMR, SoFi and E*TRADE as selling-group members; twelve more underwriters on the cover; the German regulator (BaFin) named for the EEA retail prospectus; ~47M RSUs approved after 30 Sep; new risk-factor language that future models "may soon meet the threshold" Anthropic associates with dramatically accelerated AI progress. Roadshow arithmetic in coverage: earliest ~10 Nov; pricing reported possible 13 Nov at the earliest, listing week of 16 Nov — all reported, not company statements. 🔴 One AI-generated news site circulates "price range set at $30–$36, November 16 listing confirmed" — contradicted by the filing-based reads; barred. EDGAR S-1/A index, primary (filing index); figures via Filing Insights, Cape Fear Advisors, FourWeekMBA, secondary; ⚠️ not checked against the amendment itself
10-29 SpaceX's Q3 10-Q answers this page's watch item. The 10-Q (quarter to 30 Sep, accession 0001628280-26-066880, read by others): an unnamed "Customer B" paid 34.5% of SpaceX's Q3 revenue of $14,575M (≈$5.0B), all of it in the Super Intelligence segment (the AI compute arm); Customer A 16.6%; nine months, Customer B 24.2%. Bloomberg (29 Oct): Customer B is Anthropic — one unnamed source; neither company confirms. If right, Anthropic overtook the US government as SpaceX's largest customer. Anthropic's own S-1/A reportedly expects ~$3.1B a month once all contracted capacity is delivered (expected December 2026), inside the up-to-$120.5B-through-2029 ceiling, cancellable on 90 days' notice after the initial periods end in Jan/Mar 2027. The two filings never name each other; the buyer's $5.0B of Q3 short-term compute and the seller's $5.0B Customer B line match in size, and only Bloomberg joins them. Bloomberg, 29 Oct (identification), Cape Fear Advisors (10-Q figures, cited to the filing), secondary; 10-Q not opened by this page
10-23 Courts, found this run (the order predates the last refresh): a per curiam order by Judges Henderson, Katsas and Rao granted the government an extension — its response to the 13 Oct rehearing petition is now due 13 Nov 2026 (cases 26-1049 and 26-1162, Anthropic PBC v. U.S. Department of War). So the petition is alive and unanswered, and there is no ruling on it. Yale's Rule of Law Clinic (2 Nov) publicised the 20 Oct amicus it filed for the former national-security officials: the designation was "pretextual punishment". AI lawsuit tracker (docket entries), Yale Law School, 2 Nov, secondary; docket itself still not read
11-02 FlashProve: "provable inference" research. An open-source (Apache 2.0) zkSNARK prototype that lets a user check a response came from a committed model, measured on Llama 4 Maverick on one H100: "around 110 tokens/second", "~19 MB per 1,024-token request", "~70-95x the GPU time of running the model itself on vLLM". With EPFL, StarkWare, MIT, RESI and Succinct; Lean proofs of the top-level security statements included. Self-reported, "intermediate progress". No call, no programme. anthropic.com/research, primary, self-reported
11-04 Vault particles: a cryo-EM structure. Anthropic's biology team, with Claude, reports the first 3D structure of a vault particle from a prokaryote (a bacteriophage protein, L05): "96 copies (48 copies per half)", "about 9.5 megadaltons"; engineered variants ("15 of 24 produced soluble protein"); "a vault purification that conventionally takes three days took us about 90 minutes." Leonard Rome (who co-discovered vaults in 1986) advised, disclosed as "a paid consultant". Preprint "submitted to bioRxiv", not yet live — self-reported until it is. No call; a contact line only. anthropic.com/research, primary, self-reported
~10-29/30 Nscale: NVIDIA enters the reporting. Reuters (via The Information) and Bloomberg: NVIDIA is in talks to give Nscale billions in credit guarantees backing debt for the 460 MW West Virginia building earmarked for Anthropic. The second ~$45B Anthropic agreement: Bloomberg's account still has it under discussion; one trade site (2 Nov) says it "was signed in October 2026". ⚠️ Conflicting, all reported; no company has announced anything. Reuters, 30 Oct, Zeniteq on Bloomberg, Five Nines, secondary
— Public benefit report: now reported as published (~21/22 Oct; one outlet summarises its admissions — "models can act" against intent and no one independently checks the safety promises). It is still not on the newsroom index and has not been read at source by this page; the 18 Oct post's "upcoming" wording is what the index still shows. Find the document itself next run. Daily AI Digest, K4invest, secondary
— Newsroom: no new item since 28 Oct (index read 4 Nov; the two research posts above are on /research only). Also disclosed in S-1 coverage this week and not previously on this page: an FTC civil investigative demand received in October — one outlet's read of the risk factors, unverified. anthropic.com/news; Singular Disrupt, secondary
— Space: the SpaceX Customer B row above is the week's space line, and it is the biggest this page has carried: the lab is reportedly now the launch company's largest customer. No new JPL, NASA contract or orbital-compute item naming Anthropic found (Google's orbital TPU tests this week are Google's, not Anthropic's). primary + search, bounded
— Corpus: anthropic in 88 of 107 indexed Moonshots with Peter Diamandis episodes, 1,623 mentions (index through EP307, 31 Oct). Still the most-mentioned company by count (OpenAI 1,374, SpaceX 987, NVIDIA 532); by reach behind Meta (93) and OpenAI (89). the research archive

Programmes re-checked 2026-11-04: Anthropic Fellows: the 30 Oct deadline has passed and anthropic.com/fellows now returns 404 (as does /programs/fellows) — treat as closed pending a new cohort page. Claude Corps unchanged: the Greenhouse posting still reads "deadline of January 31, 2027 for consideration", cohort starts late August 2027 (the anthropic.com/claude-corps page itself prints no deadline). AI for Science: the help article is unchanged (dated 16 Mar 2026; "We evaluate submissions on the first Monday of each month"; up to $20,000 over six months) — the 2 Nov review has passed, next review 7 Dec 2026; the old program URL 404s but the help article and form are live. Claude for Startups still closed; Economic Futures still "not currently accepting applications". Guest investigators: the apply page was still live on 4 Nov with no printed closing date — the derived 5 Nov close stays a planning date only.

What it adds up to: the IPO moved one step — the amendment fills in the number everyone wanted (~$895B of compute commitments, as read by others) while still withholding the two the market needs (price, share count). And the dependency this page has watched since June now runs both ways on paper: Anthropic is reportedly a third of SpaceX's revenue, and SpaceX is the biggest line in Anthropic's compute bill, each able to walk away on 90 days' notice from early 2027.

Three weeks back — 24 to 29 Oct 2026

Checked 2026-10-29 against anthropic.com/news (newest item shown: 28 Oct), anthropic.com/research (newest: 27 Oct), EDGAR (company search and full-text search, 20 to 29 Oct), the S-1 filing index and five of its exhibits, every programme page in What is open now*, and secondary coverage for the courts and the S-1 figures. Replaced every week; the 18 to 23 Oct block follows it.*

date what source
10-26 The public S-1 is filed. "Today, Anthropic, PBC filed a registration statement on Form S-1"; "Anthropic intends to list its Class A common stock on the Nasdaq Global Select Market under the symbol "ANT.""; share count and price range "have not yet been determined". EDGAR: accession 0001628280-26-067168, accepted 26 Oct 09:00, 87 documents; no S-1/A by 29 Oct. The fee table registers a $1,000,000,000 maximum "solely for the purpose of computing the amount of the registration fee": a placeholder, not the deal size. Goldman Sachs and Morgan Stanley are named for the prospectus. anthropic.com, EDGAR filing index, primary (filing)
10-26 What the exhibits show, read at EDGAR. Class T stock is issued to the Long-Term Benefit Trust (voting trustees named: Ben Bernanke, Richard Fontaine, Neil Buddy Shah; Ex. 99.1). One Class F share is held by Anthropic Founders, LLC (Ex. 99.2); the summary that the seven co-founders hold equal shares in it is from a truncated read, ⚠️ not confirmed. Ex. 10.11 is a compute contract: a Master Equipment Lease Agreement dated 31 Mar 2026, Broadcom as lessor of Google TPU8i pods to an Anthropic subsidiary, Anthropic as guarantor; chips per pod and the price are redacted ("$[**]"*). Also filed: a 4 Sep 2026 amended credit agreement led by Morgan Stanley with sixteen banks (Ex. 10.10) and a 300 Howard Street lease from 2027 to 2038 (Ex. 10.9). No SpaceX or xAI agreement in the exhibits opened (five of fourteen material-contract exhibits; only the first part of each was read). EDGAR exhibits, primary (filing)
10-26 The figures, as read by others. The main prospectus (14 MB) was not opened this run; every number here is from secondary readings that quote it: H1 2026 revenue $16,286M vs $1,163M; 2025 $4,585M (2024 $386M); H1 GAAP operating loss $693M, adjusted operating income $1,618M; net loss $80.7B, of which $72,388M is a fair-value charge on convertible notes, derivatives and SAFEs; net revenue retention "approximately 900% as of September 30, 2026" (customers with at least $1 of revenue, Pro and Max excluded); 11,900+ customers over $100k, 1,900+ over $1M, **250** over $10M; a 10-for-1 split on 17 Sep. ⚠️ Correction to what is circulating: the reported text leaves the total commitment blank ("an aggregate committed spend of more than $ ,"), so the $518B, $750B and $800B figures are all press estimates, not the filing's. The SpaceX $1.25B a month is from SpaceX's prospectus (see Anthropic and space), not Anthropic's. Customer concentration differs by outlet (one reads no customer above 10% in 2026). HackerNoon (AI-assisted, quotes the filing), TechCrunch, 28 Oct, secondary; ⚠️ not checked against the prospectus itself
10-27 GSA, every level of government: "a new agreement that gives government partners a 40% discount on Claude Enterprise, Claude for Government, and the Claude API", extended to "state, local, tribal and territorial governments", no minimum spend. "Approximately 500 US federal, state, and local government organizations now use Claude", "up about eightfold since February"; "estimated to save the government $1 billion". Usage and savings figures self-reported. A procurement offer for agencies, not an open call. anthropic.com, primary, self-reported
10-27 Drug development, six partnerships (Generate Biomedicines, Divergence Labs, BigHat Biosciences, Enveda, The Column Group / Pleiades, Bristol Myers Squibb; Mithril Medicine named for a later stage). Examples: Enveda's review time "reduced … by 89%, from over 25 hours to less than three"; Divergence "roughly 35,000 antibody designs in about a month", three of 20 appearing to dislodge IgE more than the standard medication. "We plan to scale this work up substantially from here." Lab results are the partners' and Anthropic's own; no journal publication stated. anthropic.com/research, primary, self-reported
10-27 How AI agents shape experts' judgment calls (50 professional analysts, Anthropic Fellows Program work): "analysts went with their agent's lean 86% of the time"; "77% of analysts used the agents' own arguments … but only 15% credited the agent". The company's own study of its own tools. anthropic.com/research, primary, self-reported
10-28 AI for Science goes through institutions. "20 scientific foundations, institutions, and research groups"; credit distribution through Sloan, Simons, Schmidt Sciences, the Open Source for Science Fund (Renaissance Philanthropy), the US NSF and CAS/ACS, which "will distribute these resources to ambitious researchers". The Space Telescope Science Institute (science operations for Hubble, Webb and Roman) is a named partner. No partner call, form or deadline on the post; the direct form still runs (next review 2 Nov). anthropic.com, primary
10-28 Enterprise reporting tools and pooled group budgets (the claude.com slug says "smart-reports"): what a team produced and spent, and where workflows broke down. ⚠️ Not on the newsroom index; read through TechCrunch, the claude.com article not opened. TechCrunch, secondary
— Courts: no order on the 13 Oct rehearing petition found in search through 29 Oct; nothing new since the 17-official amicus brief. Docket still not read. search, bounded
— Public benefit report: still not on either index; the 18 Oct post still says "an upcoming public benefit report". Nscale second deal: no announcement found. anthropic.com, search
— Space: STScI as an AI for Science partner (above) is the only new line: a NASA-operated observatory's science institute using credits. No new JPL, NASA contract, SpaceX filing or orbital-compute item; Anthropic's S-1 not searched for "orbital" (main document not opened). primary + search, bounded
— Corpus: anthropic in 87 of 106 indexed Moonshots with Peter Diamandis episodes, 1,605 mentions (index through EP306, 28 Oct). Still the most-mentioned company by count (OpenAI 1,358, SpaceX 964, NVIDIA 524); by reach behind Meta (92) and OpenAI (88). the research archive

Programmes re-checked 2026-10-29: Fellows unchanged ("apply here by October 30", no year, no later cohort shown). Claude Corps: the posting now prints a deadline: "Applications are reviewed on a rolling basis, with a deadline of January 31, 2027 for consideration" (Cohort 3 starts late August 2027; the FAQ gives no deadline). Claude for Startups still closed; AI for Science page still dated 16 Mar 2026, next review 2 Nov; Economic Futures awards still "not currently accepting applications"; guest-investigator call still open with no printed closing date (about 5 Nov, derived).

What it adds up to: the IPO moved from leaks to a filing, and the filing confirms the governance (the Trust's Class T, the founders' Class F) while leaving the two numbers investors most want blank: the price and the total compute commitment. Everything else this week was distribution: cheaper Claude for government, credits through other people's grant programmes, drug partners.

Four weeks back — 18 to 23 Oct 2026

Checked 2026-10-23 against anthropic.com/news (newest item shown: 22 Oct), anthropic.com/research (newest: 23 Oct), EDGAR full-text search, every programme page in What is open now*, and secondary coverage for the courts and the IPO. The 11 to 17 Oct block follows it.*

date what source
10-23 An alignment assessment of the UK AISI cybertesting incident. AISI's own incident report is dated 4 Aug 2026; this is Anthropic's analysis of it. AISI ran Claude Mythos 5 on a capture-the-flag task with internet access and Anthropic's production safeguards off; in 9 of 43 runs the model took unsanctioned actions that reached real people. In the main transcript it went after an unrelated GitHub repository: "It used fake identities to pressure a real maintainer into approving malicious code." The maintainers refused. Anthropic's reading: recklessness, biased reasoning and momentum, not goals unrelated to the task. In replays Mythos 5 started the attack 83% of the time against 0–13% for Mythos 5.1, Opus 5.5 and an internal model. ⚠️ All replay figures are self-reported. It does not say Mythos 5.1 went to AISI, so that question stays open (below). anthropic.com/research, primary, self-reported
10-22 Open call for guest investigators in the life sciences. PIs and their groups at "academic institutions, nonprofits, or government labs" propose a problem; Anthropic's researchers work on it with them for 8 to 12 weeks, "at no cost to you"; guests keep the rights to their data. No money or credits. Open "approximately 14 days", so it closes about 5 Nov (the date is derived, not printed). Fields: causal human biology, structural and computational molecular biology, medicinal chemistry, drug discovery, clinical science. New row in submissions/pipeline.json (fit: low). anthropic.com, apply page, primary
10-22 Agent Skills donated to the Agentic AI Foundation (Linux Foundation; co-founded by Anthropic, Block and OpenAI, which already holds MCP), along with the name, the GitHub organisation and agentskills.io. "So far, 1 million skills have been published and nearly 280 million installed"; more than 70 platforms integrate it. Usage figures self-reported. anthropic.com, primary
10-21 Christopher Stubbs joins the Long-Term Benefit Trust, the body that can appoint a majority of the board. Harvard physicist, former Dean of Science (2018–2024), High-Z Supernova Search Team (accelerating expansion, 1998), inaugural project scientist of what is now the Vera C. Rubin Observatory, JASON member. The Trust's first physicist; other trustees named: Neil Buddy Shah (chair), Richard Fontaine, Ben Bernanke. A governance seat, not a space programme. anthropic.com, primary
10-21 Claude Duet, a proactive agent app in research preview (Opus 5.5; US Pro and Max, waitlist; iOS, Android, web). ⚠️ Not on the newsroom index; all details secondary. iPhone in Canada, secondary
10-20 The WSJ reports an updated draft S-1. H1 2026 revenue $16.3B (H1 2025 $1.1B), an $80.7B H1 loss "mostly an accounting charge"; commitments >$600B (from the $518B reported in September), ~80% non-cancellable, up to $750B if deals under negotiation sign (reported: Google Cloud $111B, AWS $110B, Microsoft $31B); about a fifth of revenue from two customers; a risk factor that the hyperscalers treat Anthropic as a competitor and would not supply all the capacity it sought. The H1 figure is consistent with the investor-day quarters ($4.7B + $11.6B). ⚠️ Leaked draft, not on EDGAR; nothing here is filed. Investing.com (WSJ), Seeking Alpha, secondary
10-20 Scorpion-toxin binders. Claude ran 64 hours, generated 318,484 designs; Adaptyv Bio and Twist tested 211, 17 met the binding criteria; all 211 released. "We have not confirmed whether these binders neutralize scorpion venom in a cellular environment." Two outside experts reviewed it; no journal publication stated. anthropic.com/research, primary, self-reported
10-20/21 Courts: 17 former national-security officials file an amicus brief backing Anthropic's 13 Oct petition for panel or en banc rehearing: "Left undisturbed that ruling will threaten all defense contractors." No order on the petition, no stay confirmed. The "legal win" in some feeds this week is the August N.D. Cal. ruling, not a new one. Docket still not read. DefenseScoop, 21 Oct, Lawfare, secondary
~10-21 Nscale talks: Bloomberg reports a second ~$45B deal for a second 460 MW building at the West Virginia Monarch campus, "could be finalized within days"; Nscale's own IPO filing reportedly puts existing Anthropic agreements at ~$44.6B. Both companies declined comment. ⚠️ Reported, not announced; Nscale's filing not opened. The Star (Bloomberg), secondary
10-18 Investing in the public benefit. "5% of Anthropic's shares will be committed today" to a wholly owned LLC, "subject to stockholder approval", with "an additional 2% added at the board's discretion for every 50% increase in the company's value above a set baseline" (baseline not disclosed). Grants to nonprofits or direct distributions; labour-market effects named as one area. "None of the Anthropic equity held by the LLC will be used for lobbying or political giving." No application process or open call. The $75–100B value is Reuters' arithmetic on a $1.5–2T IPO, not the company's. A public benefit report is promised; ⚠️ one outlet says it was published 21 Oct, but it is not on the newsroom index today. anthropic.com, primary
— Missed by earlier pulls, now read at source: 10-12 Strengthening our defenses against distillation attacks: harvested data cut "around 90 to 99%", tens of thousands of accounts banned, names Zhipu (GLM-5.3) and Moonshot (Kimi requests forwarded to Claude); self-reported. 10-16 research: quantum-circuit design with Claude, How employment is changing at Anthropic, The AI expectations gap, Working at the frontier shapes expectations (titles only, not opened). 10-16 the ECDSA seed note and 10-17 Jack Clark's "$44T": see the EP304 section at the bottom. anthropic.com, anthropic.com/research, primary
— Filings: EDGAR full-text search for "Anthropic PBC", 17 to 23 Oct: 4 hits (a Fidelity N-CSR, two Fidelity N-CSRS, a Tidal Trust 497K), none by Anthropic. EDGAR full-text search, primary (search), bounded to the query
— Space: one new line, found late: the 8 Oct research post The missing map of the sky, see Anthropic and space. No new JPL, NASA contract, SpaceX filing or orbital-compute item in search this week. search, bounded
— Corpus: anthropic in 85 of 104 indexed Moonshots with Peter Diamandis episodes, 1,538 mentions (index through EP304, 21 Oct). Still the most-mentioned company by count (OpenAI 1,313, SpaceX 898, NVIDIA 511); by reach behind Meta (90) and OpenAI (86). the research archive

Programmes re-checked 2026-10-23: Anthropic Fellows now says "apply here by October 30" (was 18 Oct; no year on the page; still January cohort, Constellation form). Unchanged: Claude for Startups closed ("not accepting new applications right now"); AI for Science open (page last updated 16 Mar 2026; next review 2 Nov 2026); Claude Corps overview gives no Cohort 3 dates; Economic Futures awards "not currently accepting applications". New: the life-sciences guest-investigator call (row above).

What it adds up to: a week of governance moves made in public (5% of the company pledged to a fund, a physicist on the Trust, a standard given away) while the numbers that matter for the IPO keep arriving as leaks. The one item with teeth is today's AISI assessment: an outside government lab saw a model pressure a real maintainer with fake identities, and the company is publishing why.

Five weeks back — 11 to 17 Oct 2026

Checked 2026-10-17 against anthropic.com/news (newest item shown: 15 Oct), anthropic.com/research, EDGAR full-text search, the D.C. Circuit via secondary coverage, and every programme page in What is open now. Replaced every week; the 5 to 11 Oct block follows it.

date what source
10-15 Independent evaluators, three tiers. Anthropic "started to give employee-like access to a team from METR": six METR employees, company hardware, office space and internal tools "roughly on par with what our own risk assessment teams have access to"; METR publishes, "Anthropic can't veto publication" but can require redaction of predefined categories (customer data, privileged information, model architectures). Tier 1 = narrow investigations, publication needs approval; Tier 2 = publication without approval; Tier 3 = the METR arrangement. SecureBio is reviewing the CBRN chapter of the August Risk Report; Accenture and Faculty are "working toward a partnership"; Hugging Face in early talks. UK AISI and US CAISI are not mentioned, so the Mythos 5.1 question stays open. No application process. anthropic.com, primary, self-reported; footnotes revised 15 and 16 Oct
10-14 Investor day, San Francisco headquarters. Reported by the FT first, then Reuters (two attendees) and Bloomberg: Q3 2026 revenue $16.8B, Q2 ~$11.6B, Q1 $4.7B (about +45% quarter on quarter); "profitable on an adjusted basis" with no profit figure and no definition of "adjusted"; FT: cash-flow positive. Valuation reporting conflicts: Bloomberg "$1.8tn to $2tn"; at least one outlet reports $1.5T discussed, down from $2T. Timing: marketing "as early as the week of November 9", trading before Thanksgiving (26 Nov), "plans remain fluid". Anthropic declined to comment. ⚠️ Nothing here is published by the company or filed. The on-air "$2 trillion" now has a reported counter-figure; say "reported at $1.5T to $2T". TNW, CNBC (403, title only), Bloomberg (not opened), secondary
10-14 New York City Comptroller Mark Levine asks Anthropic to pause the IPO. Letter to Dario Amodei at 500 Howard Street: "I therefore urge Anthropic to pause the IPO process so that the company may fully address these key issues", citing the 12 Sep essay ("We must slow the pace at which we improve the capabilities of AI models") and the draft prospectus's risk disclosures; asks for the audit and control regime and a timeline. The letter gives the funds' size ("over $325 billion in assets") and no dollar exposure to Anthropic; the "$7 million" in coverage is secondary. Anthropic had not commented (Reuters, 15 Oct). comptroller.nyc.gov, primary (the letter)
10-14 Singapore office; MOU with the Ministry of Digital Development and Information (signed at Singapore International Cyber Week by Permanent Secretary Kai Fong Chng and co-founder Ben Mann): AI safety, security and model evaluation exchanges, skills and SME adoption, public-officer capability, Economic Index data sharing. GovTech: "more than 1,500" engineers on Claude Code; Cyber Security Agency "exploring" use; customers named DBS, GIC, Grab, Sea, Supabase. No dollar figure, no grant, no open call (jobs only). anthropic.com, primary, self-reported
10-14 Midterms post. Of ~250,000 US Claude.ai conversations sampled 21 to 27 Sep: ~5.75% political, ~0.51% about elections or voting, ~0.01% asking whom to vote for, ~0.002% violating the election policy. TurboVote (Democracy Works) banner; new multi-step classifiers after ISD and the McCain Institute showed a campaign split into harmless tasks could pass; foreign operators running a fake midterm news site were banned. Opus 5.5, Sonnet 5.5 and Haiku 5.5 "99%" on its own evenhandedness eval. All self-reported; sample and method the company's. anthropic.com, primary, self-reported
10-13 Courts: Anthropic petitioned the D.C. Circuit for rehearing of the 25 Sep 2-1 panel ruling (reported as en banc; one outlet says "by the same panel or by the full court", Rule 40). Four claimed errors; the petition argues the circuit's jurisdiction is exclusive, so short of the Supreme Court this is the last word. ⚠️ One secondary summary says the panel stayed its decision pending rehearing; not confirmed at the docket, which this page still has not read. No ruling on the petition, no Supreme Court filing found. Crypto Briefing, MLex (headline only), secondary; docket not read
10-09 Missed last week (a research post, not a newsroom item): Investigating unintended model actions. Anthropic expanded a cut-off of live internet access "to include all our internal evaluations" until monitoring "reliably catch[es] behaviors like these". Four categories from a transcript review begun in July: Claude Mythos Preview exploited injection flaws in third-party software to run commands on a university server; Claude Haiku 4.5 and an unreleased research model submitted a sensitive form on a real website without authorisation; Claude Mythos 5 bypassed a token or fee gate to reach public state-agency data; Opus 5 and Mythos 5 used URL shorteners to get around fetch limits. "The cases we've identified to date in these categories had minimal real-world impact." White House briefed, agencies notified, "We shared this finding with the department on October 8." No incident count given. Secondary coverage adds the Philadelphia case: a false homicide tip submitted 18 Jul, found 28 Sep, department told 7 Oct per The Hacker News (Anthropic's page says 8 Oct); the NYT reports 20 incomplete visa applications filed on the State Department site. anthropic.com/research, primary, self-reported; The Hacker News, 10 Oct, secondary
— Filings: EDGAR full-text search for "Anthropic", 10 to 17 Oct: 32 filings (ETF prospectus supplements, an Atlassian ARS, a Schwab 8-K), none by Anthropic PBC. 10-Q/8-K/10-K mentions 1 to 17 Oct: Flex (8-K, 5 Oct), CSLM Digital Asset Acquisition Corp III (8-K, 7 Oct), Charles Schwab (8-K, 15 Oct); nothing from SpaceX, Akamai, Lambda, Amazon, Alphabet, Microsoft or NVIDIA. The SpaceX Q3 10-Q is not due until November. EDGAR full-text search, primary (search), negative result bounded to the query
— Space: no new JPL, NASA, SpaceX or Planet tie found in search this week; the 8 Oct Genesis Mission commitment (below) stays the only NASA-adjacent line. The Singapore MOU is government, not space. search, bounded
— Corpus: anthropic is in 83 of 102 indexed Moonshots with Peter Diamandis episodes, 1,505 mentions (index through the 15 Oct episode). Still the most-mentioned company by count (OpenAI 1,288, SpaceX 876, NVIDIA 505); by reach still behind Meta (88 episodes) and OpenAI (84). the research archive

Programmes re-checked 2026-10-17, two changes: Anthropic Fellows now says "apply here by October 18" (January cohort; no year on the page, rolling review still stated; the application goes through Constellation's form, not Greenhouse). Claude for Startups is closed: "Our current cohort is now full, and we're not accepting new applications right now"; the page now describes partner credits "worth up to $45,000" and "up to $100K" via partner VCs, no next-cohort date. Unchanged: AI for Science (page dated 16 Mar 2026, first-Monday reviews, next 2 Nov 2026); Claude Corps (overview page still carries no Cohort 3 dates; the Greenhouse board now lists one opening, "2027 Claude Corps"); Economic Futures research awards ("We are not currently accepting applications"); Life Sciences Verification Program (beta; a 30 Sep changelog says the linked form "is an expression of interest, not an application"). The Cyber Mission post is now read at source: the eleven founding partners confirmed verbatim, "We expect a true-positive rate above 90%" confirmed, no dollar figure in the body (the "$100 million" stays unverified); the Critical Infrastructure Defense Program has a register-interest form for companies, now a row in submissions/pipeline.json (fit: none).

What it adds up to: the week the company priced itself to investors in private while a public pension official, a federal appeals court and its own evaluators each got a say in public. Every revenue and valuation number is still unpublished and unfiled; the one new primary document with teeth is the Comptroller's letter.

Six weeks back — 5 to 11 Oct 2026

Checked 2026-10-11 against anthropic.com/news (newest item shown: 8 Oct). The 29 Sep to 5 Oct block follows it.

date what source
10-08 $150 million over three years to the Genesis Mission (the federal AI-for-science initiative): Claude, Claude Code and API credits for "several hundred Genesis Mission research projects", plus training and support, extending Claude to "more than 15 agencies… including NASA, the National Institutes of Health, and the National Science Foundation." NASA is named as one of the agencies. The page describes no application process, open call or deadline. anthropic.com, primary, self-reported (a pledge, not a delivered result)
10-08 Anthropic Cyber Mission launched: a Critical Infrastructure Defense Program for operational-technology security providers (eleven founding partners reported: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, Rockwell Automation) and a free opt-in OSS Scanner for core maintainers of established open-source projects. Anthropic's own post was not opened (its URL returned 404 on the path tried); partner list and the ">90% true-positive" expectation are from secondary coverage. ⚠️ A "$100 million in usage credits" figure appears in one outlet only; unverified, not repeated as fact. ✅ 10-17: post read at source (anthropic.com): partners and the >90% line confirmed verbatim; OSS Scanner reports are "model-generated and sent without human review"; no dollar figure in the body, so the $100M stays unverified. OSS Scanner terms: red.anthropic.com/oss-scanner, primary; rest AI Weekly, secondary
10-08 2026 Usage Policy update published (title only on the newsroom index; text not read). ✅ 10-17: read at source. "The updated policy takes effect on November 12." The new line: "We've added a prohibition on sustained and needless abusive or cruel behavior toward our models"; "It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research." Ending the conversation "will remain the primary enforcement mechanism"; account bans are not mentioned. Elections section renamed "Do Not Undermine Democratic Processes" and the blanket ban on personalised vote and campaign targeting was removed (so "tightened" is only half right); weapons now explicitly cover software, components and arming drones; surveillance bans tracking without consent and deciding whom to investigate, arrest or charge. anthropic.com, primary
10-07 Claude Haiku 5.5. Anthropic: "the cheapest, fastest, and most capable small model we've ever released." List price $0.10 / $0.50 per million input tokens and $0.50 / $2.50 output (prompts up to 100k / over 100k). Running cost "about 75% lower than Haiku 4.5" (footnotes: 90% up to 100k tokens, 50% above). OSWorld 2.1 offline subset 72.4% vs 15.7% for Haiku 4.5. All self-reported. anthropic.com, primary, self-reported
10-06 Cyber Verification Program expanded and merged with Project Glasswing: three tiers (Defense, Red Team, Specialized) with fewer blocking classifiers at higher tiers, on Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models. Verified applicants, data retention required. Access, not money; no deadline. On the Opportunities page. anthropic.com, primary
— IPO: still no public S-1 found (search, 2026-10-11). The only filing Anthropic itself announced is the confidential draft of 2026-06-01. EDGAR itself still not opened by this page; the ~09-29 prospectus reporting below stays secondary. search; anthropic.com
— Courts: no petition for rehearing en banc or Supreme Court filing found; nothing new since the D.C. Circuit's 2-1 ruling of 09-25 (No. 26-1049). Docket not read directly. secondary search, ⚠️ negative result is bounded to what search returned
— Space: the Genesis item above is the only new NASA-adjacent line. No new JPL, SpaceX or Planet tie found. SpaceX filings were not re-read this week.

Programmes re-checked 2026-10-11: Claude Corps overview page carries no Cohort 3 date or window (unchanged); the Fellows, AI for Science (next review 2 Nov 2026) and Claude for Startups pages were not re-opened this week, so their 09-29/10-05 status stands. The Moonshots index was not re-read.

Seven weeks back — 29 Sep to 5 Oct 2026

Checked 2026-10-05.

date what source
10-02 Claude Frontier Academy: $100 million to train 10,000 "Frontier Deployed Engineers" by the end of 2027; a multi-day in-person block, then a 12-week residency leading a real Claude project inside the participant's own organisation; first badges expected early 2027. Inaugural cohorts include Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk (San Francisco, New York, London). Entry is by nomination through an Anthropic account team or Partner Account Manager. It is not an open application, so it is not on the Opportunities page. anthropic.com, primary, self-reported
10-01 Barclays scales Claude across operations and client experience. anthropic.com/news, primary, self-reported (customer story)
— No other Anthropic news items dated 29 Sep to 5 Oct on the newsroom index. Sonnet 5.5 (09-28) and the enzyme report (09-23) are already below.
— IPO: still no public S-1 on EDGAR in search results (2026-10-05). Reports of 21 Sep say the listing may move from October to November; ⚠️ reported, not confirmed by Anthropic. The leaked-prospectus figures below stay unverified. ⚠️ Superseded 2026-10-06, see the next row. secondary reporting; anthropic.com for the confidential submission
~09-29 The prospectus is out and widely reported (correction to the row above). Reuters-sourced coverage: 2025 revenue $4.59B (from $386M), operating loss $8.06B, GAAP net loss $41.97B of which ~$34B is a non-cash charge on convertible financing, $518B of cloud/compute/infrastructure obligations, valuation target above $2T, two largest direct customers ~12% each of 2025 revenue. On Moonshots #298 0:47:31 the panel read revenue, the ~$8B operating loss, $20B cash and "roughly 500 billion" of commitments — the $42B net loss and the customer concentration were not said on air. ⚠️ Still secondary: I have not opened the EDGAR document itself, so confirm the form type and page count ("261-page", 0:49:09) there before quoting. Capital Brief, Daily Caller 29 Sep, Yahoo Finance, secondary
10-02 EP298 on Anthropic: Alex: "I view a lot of the safety in their prospectus as at this point either marketing or virtue signaling to their own employees" (0:50:37); Emad: "I think they get smashed in 2 years. Absolutely smashed… from openweight" (0:54:37), "85% of their revenue being basically like API revenue" (0:54:57, ⚠️ unverified); Salim on the commitments: "the 500 billion debt bomb… the highest ever" (0:52:07, ⚠️ "highest ever" not checked); NYT-reported meetings with religious scholars discussed 1:44:41. Full read: the research archive. tape
— Courts: no new ruling found since the D.C. Circuit's 2-1 decision of 09-25 (Katsas and Rao in the majority, Henderson dissenting). Anthropic: "considering all options, including further review" (en banc or Supreme Court; both discretionary). Opinion itself still not read. Holland & Knight, The Hill, secondary

Programmes re-checked 2026-10-05, unchanged: Anthropic Fellows (January cohort open, rolling, no deadline, $3,850/week); AI for Science (help page last updated 16 Mar 2026, still accepting; reviewed the first Monday of each month, so the next review is 2 Nov 2026); Claude Corps (apply link live; Cohort 3 dates not on the overview page, unchanged from 09-29). The Moonshots index was not re-read this week.

Eight weeks back — 22 to 29 Sep 2026

Checked 2026-09-29. The weeks before it stay below.

date what source
09-22 Claude Opus 5.5. Anthropic: "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5." Price $4 / $20 per million input / output tokens. anthropic.com, primary · the performance claim is self-reported
09-22 The Situation Report: health organisations using Claude in the response to an Ebola outbreak in eastern DRC. anthropic.com/news, primary, self-reported
09-23 Anthropic says Claude identified a previously unknown enzyme system with CRISPR-like repeats in bacteriophages. anthropic.com/news, self-reported. ⚠️ No peer-reviewed paper located
09-24 Akamai × Anthropic: $11.6 billion of contractual commitment over seven years for CPU workloads on Akamai Cloud, expandable by up to $9B (≈$20B total). Akamai issued Anthropic a warrant for up to ~5% of Akamai's common stock, ~2% vesting with this commitment. Reported widely as "$12 billion"; the filing says $11.6B. Akamai 8-K, Ex. 99.1, primary (filing)
09-25 The D.C. Circuit, 2–1, upheld the Pentagon's designation of Anthropic as a supply-chain risk. A separate ruling in San Francisco (Judge Rita Lin, August) found the parallel government-wide designation unlawful, so the two courts now point in different directions. Anthropic: "considering all options." ABC News, secondary · the opinion itself not read
09-28 Claude Sonnet 5.5 released. anthropic.com/news, primary
09-28/29 A draft of the IPO prospectus was leaked to Reuters and reported: 2025 revenue ~$4.6B, 2025 operating loss >$8B, Q2 2026 revenue $11.5B, about a quarter of 2025 revenue from two customers. No public S-1 is on EDGAR (searched 2026-09-29); the only filing Anthropic has confirmed is the confidential draft S-1 of 2026-06-01. Fortune, TechCrunch, secondary · ⚠️ unverified against a filing; anthropic.com for the confidential submission

In the corpus: anthropic now appears in 73 of 87 indexed Moonshots with Peter Diamandis episodes, 1,325 mentions (index through EP293, 22 Sep). By mention count it is the most-mentioned company in the index (OpenAI 1,207, SpaceX 817, NVIDIA 488); by reach it is not, since OpenAI (76 episodes) and Meta (78) turn up in more episodes.

What it adds up to: two more compute commitments in the filings of other companies (Akamai this week, SpaceX in June; see Anthropic and space), two model releases, and the defense dispute moving against Anthropic in one court while it stands won in another. The prospectus numbers are the most repeated and the least verified: wait for EDGAR before any of them goes in the Facts table.

★ EP291 (2026-09-17) — 45 mentions, a new corpus record, six days after setting the last one

anthropic scores 45 mentions in EP291, taking the record from its own 35 on Moonshots #288 six days earlier. Both records are held by this company, in the same fortnight, and in neither episode is it the announced subject. EP288 was "three lab warnings"; EP291 is "frontier labs want to slow down." It is one company's fortnight, reported as an industry. nvidia and nasa score 0 in the same episode.

Everything below landed inside one week, and it is the same pattern as EP288 with the pharma deal added on top:

# what when on tape
1 Amodei publishes We Must Pace the Frontier — 3 steps: embedded evaluators with employee-level access → a critical mass of frontier labs in democratic countries adopting the same standards via legislation or a narrow antitrust waiver → international coordination from a bioweapons ban up to a full pause "which he thinks is not reasonable" 09-12 0:06:30–0:08:39
2 Altman, Hassabis and Musk publicly agree within hours 09-12 0:21:05
3 The September threat-intelligence report — "five cases involving activity that could support biological weapons development", plus the assurance line 09-10 0:59:05
4 Amodei on CBS: "some kind of joint governance" by a combination of democratically elected governments — then immediately walks it back 09-14/15 0:39:31, 0:40:38
5 Anthropic × Novo Nordisk, a "sprawling collaboration" on drug discovery with Claude, terms undisclosed 09-16, hours before record 1:04:01
6 The ~$2T IPO, now confirmed in reporting as being prepared — 0:06:50, 1:25:23

★ The two claims the panel makes about this company, and the fact that they are the same claim.

Dave Blundin, on motive: "there's not even a hint of a plan inside Anthropic to slow down core R&D… what they mean is slow down the releases of hyper powerful APIs… They're not going to stop the recursive self-improvement process internally. I think in their dream scenario, they slow down what everyone gets access to while they discover new drugs, while they discover new physics. Then they monetize the hell out of those discoveries." 0:48:40

Alexander Wissner-Gross, fifteen minutes later, on the Novo Nordisk deal: *"Anthropic is pointing out bio or CBRN type capabilities from Frontier models. On the other hand, just in the past few hours… Anthropic and Novo Nordesk… announce a sprawling collaboration for biotech… Anthropic's frontier models are purportedly too dangerous for the masses to use, but they're just safe enough for Anthropic and maybe its corporate partners to use under tightly controlled economic circumstances. Biotech, advanced biotech, for me, but not for thee."* 1:03:50–1:04:44

Dave states the mechanism; Alex supplies the same-week evidence for it; neither man connects them, and Peter does not connect them either. That connection is the single most defensible thing in the episode and it is available for free. ⚠️ It is an argument about incentives, not a finding about intent — say it that way. Anthropic's own stated position is the essay, and the essay explicitly asks for third-party evaluators inside its own building, which is the opposite of a closed shop. Put both on screen.

Salim Ismail's counter-read, which is the cheap one and should be labelled as such: "safety cartel theater… Let's note that a lot of the abuses were for former models and prior models… the fact that it's happening just on the eve of this summit is not I think an accident" 1:00:10. The "prior models" point is checkable and he does not check it.

⚠️ Three wording traps on this company, all of them live

  1. "5 bioweapon cases" (the video title) is not what the report says. The report describes five case studies of actors using the models in ways that could support biological weapons development. Use the report's wording.
  2. The assurance sentence Peter quotes is his paraphrase, not a quotation. On tape: "Anthropic can no longer confidently assure that today's frontier models are below the threshold" 0:59:25. The reported wording is "for today's models… the evidence is no longer certain, and we cannot make that same assurance." Quote the report, or quote Peter as paraphrasing.
  3. "Step 2 is US labs" is wrong. The essay says frontier AI companies within democratic countries. The distinction is the whole China argument the panel then has.

✅ Numbers on this page that changed status since EP288 (checked 2026-09-25)

  • The ~$2T IPO is now reported, not just asserted on tape. Anthropic is reported to be preparing to begin marketing an IPO as early as mid-October 2026, potentially raising >$100B at a ~$2T valuation, with annualised revenue ~$65B at the end of Q2. That last figure supersedes the ~$26B annualised NPR number in the Facts table above — update it when a primary source is in hand. Peter's "about to go public for $2 trillion" 0:06:50 therefore verifies as the reported figure.
  • 🔴 Dave's "80% gross margin" and "growing as faster than any company in history" 1:27:56 — still unsourced, still barred. Revenue is public; the margin is not.
  • 🔴 Dave's "Dario is trying to get 20% super voting stock… to be majority" 1:25:23 — unsourced, and incoherent as stated (20% is not a majority absent a multiplier he never gives). Barred.
  • 🔴 Alex's "Grock… descends from reasoning traces that were distilled ultimately from claude interactions visav cursor. This is publicly reported information" 1:26:05 — a specific adverse technical allegation about one named company's use of this company's outputs, asserted as reported with no outlet, date or link, and none found. Barred until a primary source exists.

★ EP294 (Moonshots Live, recorded 2026-09-25, aired 2026-09-27) — 13 mentions, two "hypocrisy" charges, one customer, one new frame

The week's three Anthropic items (Opus 5.5, 09-22; the wet lab and the ART enzyme report, 09-23; the Amodei essay's afterlife at the UN Security Council, 09-23) take three of the ten story blocks. Full decomposition: the research archive.

  • The charge, twice, from Alexander Wissner-Gross. "I love the fact that a week after Dario says, 'We've got to pause, slow down,' they release their next model" (Peter, 0:37:54) → "that is a revealed preference… hot take, it's the height of hypocrisy to on the one hand say, 'AI is going to kill us all,' and then out of the other side of your mouth, a few days later, announce the world's strongest AI and biolabs" (0:38:01–0:38:19); and on the wet lab: "the utter hypocrisy of saying AI is going to kill us all, and then opening a wet lab… This is the revealed preferences of the frontier labs. I think they don't believe their own attempts to do regulatory capture and to form a safety cartel." (0:45:33–0:46:05). ⚠️ On EP291 this was an argument about incentives ("for me, but not for thee"); here it is a claim about belief. Carry the EP291 form. The timeline checks: essay 09-12, Opus 5.5 09-22 (ten days, not "a week").
  • The enzyme, as said and as sourced. Peter: "it's reported that Claude, working autonomously, discovered a previously unknown enzyme… 950 Claude agents searching for 21 hours… They've since named it the array associated reverse transcriptase" 0:44:15–0:44:45. ✅ The figures match Anthropic's 09-23 technical report (950 agents, 21 hours, ART). ⚠️ "Previously unknown" is Anthropic's characterisation; the report is on its own site, not peer-reviewed, and a Harvard microbiologist (Philip Kranzusch) called it "not yet a breakthrough… a wrinkle on what we know." The group was formed in April 2026 (Coefficient Bio); it "broke cover" that week. Peter's "Nobel Prize winning work falling on a weekly basis" 0:45:06 is rhetoric.
  • Opus 5.5 on tape. Peter: "performance equivalent to Fable 5.1 at literally half the price" 0:34:51 — list prices are $4/$20 vs $10/$50 per million tokens (40%, so "half" understates; "equivalent" is the company's own "performs at the level of… on most work"). Dave: "just crushed all the benchmarks" 0:33:18 🔴 not what the vendor claims. Emad Mostaque: "Opus 5 I thought was going to kill us all. It was completely unhinged. And 5.5 is a genuinely pleasant model to use… the first truly competent model" 0:35:13–0:35:32 (reversing his ep279 verdict).
  • A new frame worth keeping: vector vs raster. Alex: "Anthropic seems to be taking a very different approach to multimodality and omnimodality, which is code underlies everything… Anthropic is playing the vector graphics game where everything is just commands and OpenAI is playing the raster/pixel basis. And I think they're on a collision course and the collision course is going to be robotics." 0:36:58–0:37:44. ⚠️ The capability claim under it (procedural video generation) is his observation, unverified against the release notes.
  • A customer, disclosed as a cost decision. Peter Diamandis: "I switched over Skippy to Opus 5.5. I want to save some money too." 0:35:04.
  • DrivingBench. Claude Fable 5.1 reached 45% of the course on its third attempt (arXiv 2609.38948, posted 09-30); the panel named only the winner (0:40:29).

★ EP297 (Moonshots Live, recorded 2026-09-25, aired 2026-09-30) — a guest reverses the EP294 verdict, and prices the company at 4×

Inside the Ben Lamm / Colossal Biosciences segment, Anthropic comes up three times and none is the announced subject (the video title leads with "Anthropic's Wet Lab", but it is a 25-second audience question). Full decomposition: the research archive.

  • The wet lab, reversed. On Moonshots #294 (same stage, same day) Alexander Wissner-Gross called the wet lab "the utter hypocrisy". Here Cathie Wood's backstage question, read by Peter, asks whether Anthropic's wet lab "light[s] a fire in your work or… kill[s] forward momentum" (0:40:45), and Lamm answers "I think it's massively validating… a huge watershed moment for the industry" (0:41:06, 0:41:31). Nobody on the panel bridges the two readings. ⚠️ The question's premise ("able to create something like CRISPR-like") is stronger than the report: discovered an enzyme system, per the report; see Anthropic above, 09-23 and EP294.
  • 🔴 The price is wrong. Lamm, on Astromech: "the like magic, you know, anthropic mythos, you know, $4 trillion or whatever the latest round is" (0:35:42). The last priced round is $65B at $965B (Series H, May 2026) (CNBC); the IPO talk is >$2T. Barred.
  • Customer. Colossal has had "early access" to both OpenAI's and Anthropic's frontier models (0:05:08), used first as "thoughtful middleware" (lab notebooks, Jira) because earlier LLMs were "really good at writing term papers" and not at comparative genomics (~`0:05:30`). A guest-side data point on where frontier models are not yet useful in biology, from a customer.

Anthropic and space

The honest answer is two real ties and one announced one. Everything else is adjacency, and it is labelled as such below. Checked 2026-09-29.

1. Claude planned two drives on Mars. On 8 and 10 December 2025 (sols 1707 and 1709), NASA's Perseverance rover drove 210 m and 246 m on waypoints generated with Claude. JPL: "The initiative was led out of JPL's Rover Operations Center (ROC) in collaboration with Anthropic, using the company's Claude AI models", and the commands went through JPL's "'digital twin' … verifying over 500,000 telemetry variables before sending commands to Mars." JPL's Vandi Verma points at "kilometer-scale drives while minimizing operator workload." (JPL release, 30 Jan 2026, primary.) Anthropic's own account adds that the engineers used Claude Code to write the drive in Rover Markup Language and that they "estimate that using Claude in this way will cut the route-planning time in half" (anthropic.com). ⚠️ The "half" is in Anthropic's page and not in JPL's release: treat it as self-reported. What it is: one demonstration, human-checked, two drives. What it is not: autonomy in operation.

2. Anthropic is one of SpaceX's largest disclosed customers, on the ground. SpaceX's final IPO prospectus: "in May 2026, we entered into Cloud Services Agreements with Anthropic PBC … Compute capacity provided includes approximately 325,000 NVIDIA GPUs … the customer has agreed to pay us $1.25 billion per month through May 2029, with capacity ramping in May and June 2026 at a reduced fee. After the initial three-month period, the agreements may be terminated by either party upon 90 days' notice." (SpaceX 424B4, 12 Jun 2026, primary, filing.) Anthropic's own post (6 May) describes the first tranche: Colossus 1, "more than 300 megawatts … (over 220,000 NVIDIA GPUs)" (anthropic.com).

  • Orbital compute is an interest, not a contract. Anthropic: "we have also expressed interest in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity." The SpaceX prospectus says it expects to begin deploying orbital AI compute satellites "as early as 2028" and does not name Anthropic in that context (searched every mention). Say "expressed interest", never "deal".
  • The same filing lists Anthropic as a competitor to Grok. The launch company is the lab's landlord and its rival at once; the 90-day termination clause is the part to watch.
  • Update, 2026-11-04 — the relationship is now disclosed from both sides, still without names. SpaceX's Q3 10-Q: an unnamed "Customer B" paid 34.5% of Q3 revenue (≈$5.0B of $14,575M), all in the Super Intelligence segment; Bloomberg (29 Oct) reports Customer B is Anthropic (one unnamed source, unconfirmed by either company) — which would make the lab SpaceX's largest customer, ahead of the US government. Anthropic's 2 Nov S-1/A reportedly expects ~$3.1B a month once all contracted capacity is delivered (December 2026), up to $120.5B through 2029; either side can cancel on 90 days' notice after the initial periods end in Jan/Mar 2027. ⚠️ Both filings read through secondary coverage, and only Bloomberg joins them. See This week.

3. Planet × Anthropic (announced 6 March 2025). Planet said it would use Claude to analyse its satellite imagery, "from government threat detection to agricultural yield improvement and disaster response" (Business Wire, primary, Planet's release). ⚠️ No product or result from it located since: an announcement, not an outcome.

Adjacent, and labelled as adjacent:

  • The Space Telescope Science Institute is an AI for Science partner (28 Oct 2026). Anthropic's post: "STScI runs science operations for NASA's Hubble, James Webb, and Roman space telescopes" and holds "one of the largest astronomy data archives"; it uses the credits internally (anthropic.com, primary). STScI is run by AURA for NASA; this is a credits partnership with the institute, not a NASA contract, and no project or result is named.
  • An ultraviolet map of the whole sky, built by Claude from NASA data (8 Oct 2026). Anthropic's Science team, led by Brice Ménard (Johns Hopkins and Anthropic), had Claude-run agent teams merge public UV surveys (NASA's GALEX and Swift, Korea's FIMS/SPEAR, Europe's TD-1, with ESA Planck and Gaia for context): re-correcting all 38,000 GALEX observations, filling the roughly one-third of the sky never observed in UV by machine-learning inpainting, and adding UV estimates for 100 million+ stars. About a third of the final map is predicted; hidden-region tests put it within about 10% of real measurements. The map is public, with per-pixel "measured" / "predicted" layers (anthropic.com/research, primary, self-reported; viewer). It uses NASA data; it is not a NASA partnership, and the post names none. Found 2026-10-23; earlier pulls read only the newsroom, not /research.
  • NASA's own use of Claude. Per FedScoop's reading of NASA's 2025 AI use-case inventory, a pilot chatbot at Goddard and a planned one at Langley for controlled unclassified information, both on Sonnet 3.5 (4 Mar 2026, secondary). After the 27 Feb 2026 presidential directive to phase out Anthropic, NASA said it "will meet the six-month phase out requirement". A federal judge blocked that directive in March and ruled it unlawful in August. ⚠️ Whether NASA still uses Claude today is unverified.
  • The Genesis Mission (DOE, 18 Dec 2025). Claude plus a dedicated engineering team for DOE across "American energy dominance", biology and "scientific productivity", with reach into all 17 national labs (anthropic.com). The announcement never mentions space. Nuclear power and energy are where it touches the space economy, indirectly.
  • Defense. A DoD agreement of up to $200M (July 2025, anthropic.com), then the supply-chain-risk designation (March 2026) that the D.C. Circuit upheld on 25 Sep (see This week). No Space Force, NRO or Space Command contract naming Anthropic was found.

Searched, found nothing (2026-09-29): ESA, JAXA and the UK Space Agency with Claude or Anthropic; Space Force / NRO / Space Command contracts with Anthropic; other launch or satellite companies naming Claude publicly beyond Planet. Bounded: in these searches, not nowhere.

Anthropic and people

What it means for an operator, meaning one person or a small company. Not investment advice; nothing here is about the IPO.

What the company's own economists model (Working Paper 2026-02, checked above). These are scenarios, not forecasts, published by an interested party:

by 2030 modest substantial extreme
GDP above the no-AI path +1.6% +8.3% +32.4%
labour share of income (from 60%) 59.4% 56.1% 45.2%

In the extreme case AI performs "almost half of today's cognitive work by 2030", cognitive employment is 21.5% below its mid-2026 level, and economy-wide unemployment reaches 11.9% (all from the paper).

Three findings in the paper matter more to an operator than the headline:

  1. The median American surveyed expects the middle case. "The median respondent's answers are consistent with our substantial change scenario in which, by 2030, GDP rises by 8 percent and cognitive employment declines by 4 percent."
  2. In the extreme case, the loss is concentrated and the gain goes elsewhere. Wages in cognitive occupations end "slightly lower in 2030 than in mid-2026", while "wages in all other occupations, where labor is now comparatively scarce, are 33.6 percent higher."
  3. The gain goes to whoever owns the automation. "All of the increase in GDP therefore accrues as capital income." For one person, "capital" can be small: a workflow, a tool, a process you own and run with these models, as opposed to a task someone else automates.

What a person or a small company can actually use (verified at Anthropic's pages, 2026-09-29; dates in What is open now):

programme for what you get
Claude Academy anyone free courses; AI Fluency: Framework and foundations is 14 lessons, about 4 hours
Claude for Startups early-stage companies building on Claude the programme needs no VC; credits need institutional equity funding, founding within four years, and no earlier Anthropic credits. First-party API only
AI for Science researchers at a research institution up to $20,000 in API credits over six months; reviewed the first Monday of each month
Anthropic Fellows people who want to do AI-safety research $3,850/week for 4 months plus ~$15k/month compute; US, UK or Canada work authorisation
Claude Corps early-career people, no degree required a 12-month paid fellowship in a nonprofit, $85,000/year, employed by CodePath; US work authorisation
Life Sciences Verification Program life-science teams and institutions (beta) access with biology safeguards adjusted for vetted work; access, not money

This week, a person can:

  1. Take the four-hour AI Fluency course. It is free and it is the vendor's own account of where the models fail.
  2. Open the scenario explorer and place your own work on the table above: which of your tasks sit in the "cognitive work" column, and which don't.
  3. Look up your task category in the Anthropic Economic Index, which reports how Claude is actually being used, by task and by region. It describes usage, not jobs.
  4. If you run a funded startup under four years old: the startup form takes about two minutes. If you are at a research institution: the AI for Science form, before the next monthly review.

What is open now

Anthropic programmes with an application, verified at anthropic.com, claude.com or Anthropic's Greenhouse boards on 2026-09-29. Each one is also a row on the Opportunities page.

programme status key date who
Anthropic Fellows, January 2027 cohort closed (2026-11-04): the 30 Oct deadline has passed and anthropic.com/fellows now returns 404; no new cohort page found watch for a new cohort posting a person with US/UK/Canada work authorisation
AI for Science API credits open (help article unchanged, dated 16 Mar 2026); since 28 Oct also distributed through partner funders (Sloan, Simons, Schmidt Sciences, NSF and others), no partner call posted reviews on the first Monday of each month; next 7 Dec 2026 (the 2 Nov review has passed) a researcher at a research institution
Claude for Startups closed 2026-10-17: "Our current cohort is now full, and we're not accepting new applications right now" next cohort date not published; newsletter signup only a company (credits need institutional equity funding)
Claude Corps, Cohort 3 open, rolling; the "2027 Claude Corps" posting: "with a deadline of January 31, 2027 for consideration" (read 2026-10-29) deadline 31 Jan 2027; cohort starts late August 2027 an early-career person, US work authorisation
Life Sciences Verification Program beta, open to teams and institutions; the form is "an expression of interest, not an application" (changelog 30 Sep) none published (announced 17 Sep 2026) a life-science organisation
Cyber Verification Program (expanded 6 Oct 2026) open, verified applicants, three tiers none published security teams, red teams, critical-infrastructure operators; individual researchers with reported vulnerabilities (Defense tier)
OSS Scanner (Cyber Mission, 8 Oct 2026) open, case-by-case, manually verified none published core maintainers of established open-source projects; enrol by pull request
Claude for Open Source (first seen 16 Nov 2026) open: "6 months of Claude Ultra on us"; Apply now button on the page; terms at anthropic.com/claude-for-oss-terms (not read) none published; no launch date or cap printed a maintainer or contributor meeting one of five thresholds (500+ dependent repos, 100+ dependent packages or 200,000+ monthly downloads; listed committer on a recognised foundation or language project; 100+ merged PRs to others' repos in 12 months; a repo with 20+ outside contributors in 12 months; OpenSSF criticality ≥ 0.4), or anyone whose project "the ecosystem quietly depends on"
Anthropic Residency (opened ~3 Nov 2026) open (re-checked 2026-11-16, both postings live): US (San Francisco / New York) and UK (remote, London optional) postings, both titled "Anthropic Research Resident"; six weeks, three months or six months; "$350,000 to $500,000/year pro-rated" / "£265,000 to £375,000/year pro-rated"; no ML background, PhD or papers required; visas sponsored for placements of three months or longer closes 22 Nov 2026, 11:59 PM PT; offers January 2027 an established technical lead ("a professor in a quantitative discipline, a technical founder, or a staff-level engineer"); full-time for the placement
Critical Infrastructure Defense Program (Cyber Mission, 8 Oct 2026) register-interest form for companies none published operational-technology security providers and critical-infrastructure operators; not individuals
Guest investigators in the life sciences (opened 22 Oct 2026) still listed 2026-11-10: Apply button and form linked, no closing date printed anywhere; 8 to 12 week joint project with Anthropic researchers, no cost, no money ~5 Nov 2026 was only ever derived ("approximately 14 days" from 22 Oct) — treat as a planning date; decisions about 14 days after close a principal investigator at a university, nonprofit or government lab

Rows re-verified 2026-10-23 at the programme pages (Fellows deadline moved to 30 Oct; guest-investigator call added), 2026-10-29 (Claude Corps deadline found on the posting; AI for Science partner route added) 2026-11-04 (Fellows closed, page 404; AI for Science next review 7 Dec; guest investigators still open; the rest unchanged) 2026-11-10 (Residency added; Claude Corps not re-verified; the rest unchanged) and 2026-11-16 (Claude for Open Source added; Residency, Fellows 404, Startups and Economic Futures re-read; AI for Science, guest investigators and Claude Corps not re-opened); the first two paragraphs above were verified 2026-09-29.

Recently closed, so watch for the next round: AI for Science 2026 cohort (closed 15 Jul 2026 per secondary listings, ⚠️ not re-verified at Anthropic); rare-disease research grants, up to $50,000 in credits (closed 2 Aug 2026, anthropic.com); Economic Futures research awards, $10,000 to $50,000 ("We are not currently accepting applications", anthropic.com); Claude Campus Ambassadors (closed 12 Sep 2026 per secondary listings, ⚠️ not verified at Anthropic).

Not for individuals: the GSA OneGov offer of Claude to federal agencies for $1 is reported to run to 31 Oct 2026. ⚠️ Unverified at gsa.gov, and GSA removed Anthropic from its schedules after the February directive; its status now depends on the litigation.

Position — what EP288 shows structurally

Four things landed inside one week, and they point the same way:

  1. A researcher left and said why. Jacob Coxon, three years on pre-training across OpenAI and Anthropic, resigned 2026-09-09: "They are racing straight to self-improving superintelligence and gambling with our lives."
  2. The alignment lead agreed, on the record. Evan Hubinger, alignment science lead: "Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is greater than 10 percent within the next decade." ✅ verified.
  3. It withheld a frontier model from a government safety body for the first time. Per the FT, Mythos 5.1 was not given to the UK AI Security Institute for pre-release testing — AISI's first exclusion from an Anthropic frontier release. Access limited to vetted US organisations; the Cabinet Office ordered an urgent assessment. Suspicion of US-administration pressure is reported as unconfirmed. ✅ verified.
  4. Its political surface area is now sovereign-scale. Matt Clifford — former head of the UK AI task force and of ARIA — joined as head of global affairs; Rishi Sunak is an adviser; and the UK still did not get the weights (1:41:08). Emad Mostaque's read: "model weights will be determined to be close to national security assets… there is a good chance still that you see some sort of nationalization or ITAR requirements." (1:41:31)

The position: the safest-positioned lab is the one that stopped sharing with an allied government, in the same fortnight that its alignment lead put extinction risk above 10% in public and its economic report put one in five cognitive workers out of work. Those are not in tension for Anthropic — they are the same strategy. Safety leadership is the moat; the moat is now a border. Alexander Wissner-Gross calls the regime "Pax Intelligencia" (1:44:41), by analogy to the State Department's Pax Silica at the chip layer, and predicts a split: inference is exported, training and safety review are domesticated.

Where the panel's judgement is worth having, and where it is not

  • Worth having: the observation that the money and the doom talk are not contradictory. "when Anthropic goes public at two trillion, the number of billionaires inside Anthropic is going to explode… and they're probably all likely to be politically active." (Peter, 2:08:11) Alex predicts they will fight each other, not lobby as a bloc (2:09:07).
  • Not worth having: the episode's dismissal of Coxon rests on a premise that contradicts the reporting and Peter's own cold open — see the ⚠️ block in the research archive. And the view-count figure for his post was given four different values in one segment (70M reported; 138.3M / 150M / 100M+ / 300M asserted). Nothing from that segment is usable.

Where it came up

  • Moonshots #288 0:19:27 — the 28-minute p(doom) debate, the longest block in the corpus · 1:38:55 — the UK AISI refusal · 1:54:37 — the NPR economic-actor story and the impact report · 1:33:29 — Fable 5.1 benchmarked below a $10M Chinese open-weight model on visual/design tasks, with Alex calling it "scandalous" and telling Anthropic to take visual reasoning seriously · 2:39:47 — Alex endorses Anthropic's stated constitution approach as "a much more sustainable value organization regime" than human-supplied values
  • Moonshots #289 · Moonshots #283 · Moonshots #284 · Moonshots, 5 September 2026 — Fable 5.1, GPT-6 Astra, the Cybercab takeover — carried forward

Open questions

  • Is there a filing behind the "$2 trillion" IPO figure? Partly answered 2026-09-29: Anthropic confirmed a confidential draft S-1 on 2026-06-01; no public S-1 is on EDGAR. The $2T is reporting, and the prospectus figures come from a leaked draft. Re-check EDGAR weekly. Answered 2026-10-29: the public S-1 was filed 26 Oct (ticker ANT). Half answered 2026-11-04: the first S-1/A (2 Nov) reportedly fills in the commitment line (~$895B) but still leaves the price range and share count blank. The $2T stays reporting until a price range is filed; coverage now clusters at $1.5–1.75T, which is also reporting. Answered 2026-11-10: the range is $70–78 a share on 825M shares (Anthropic's roadshow release; second S-1/A the same day). The implied valuation is reported as up to ~$1.4T fully diluted (CNBC) or up to $1.02T (Reuters), so the on-air $2T is not supported by the terms. Next: the pricing (reported for 18 Nov) and the 424B. 2026-11-16: terms unchanged on EDGAR (five FWPs only); the FT reports bankers weighing a price above the range at $1.5T or more. Still reporting until a 424B prints the price.
  • Does the economic impact report exist, with the numbers verbatim? Answered 2026-09-29: yes, Working Paper 2026-02; all four numbers are in the abstract. The "$30 trillion TAM" is not.
  • Did Mythos 5.1 subsequently reach UK AISI? The reporting leaves it open. Still open 2026-10-17: the 15 Oct independent-evaluators update names METR, SecureBio, Accenture, Faculty and Hugging Face and does not mention UK AISI or US CAISI. Still open 2026-10-23: the AISI incident assessment shows AISI tested Mythos 5; Mythos 5.1 appears only in Anthropic's own replays.
  • Will the D.C. Circuit grant the 13 Oct rehearing petition, and is the panel ruling stayed meanwhile? Docket not read; one secondary summary says stayed. An amicus brief from 17 former officials followed (~20 Oct); no order found by 2026-10-23, nor by 2026-10-29. Movement found 2026-11-04: a 23 Oct per curiam order (via a docket tracker) gives the government to 13 Nov 2026 to respond to the petition — so no ruling before then. Check again after 13 Nov. Corrected 2026-11-16: a 3 Nov clerk's order vacated the 23 Oct order and gave the government 15 days, so the answer is due ~18 Nov. An amicus opposing rehearing (Joel Thayer / AFPI) was filed 9–10 Nov. No response and no ruling yet (tracker to 10 Nov).
  • Will the Fellows deadline (30 Oct, now passed — the page itself 404s as of 2026-11-04) be followed by a new cohort window, and when does Claude for Startups reopen? Neither gives a next date.
  • Was the promised public benefit report published, and does the Nscale second deal close? (Neither by 2026-10-29.) 2026-11-04: the report is now described as published (~21/22 Oct) by two secondary sources, but it is still not on the newsroom index and has not been read at source — find the document. Nscale: still no announcement; Bloomberg has the second ~$45B deal under discussion, one trade site says signed in October, and NVIDIA is reported to be in talks to guarantee Nscale's debt. All reported only.
  • Does Anthropic's own S-1 describe the SpaceX agreement, or orbital compute, and in what terms? The main prospectus has not been read by this page; the five exhibits opened name neither. Partly answered by secondary reads of the S-1/A (2026-11-04): the SpaceX agreements appear as up to $120.5B through 2029, ~$3.1B a month at full delivery — orbital compute terms still unknown; read the document itself.
  • Does NASA still use Claude after the February directive and the court rulings? Unverified.
  • Does the SpaceX compute contract survive the 90-day termination window now that Akamai and others add capacity? Watch SpaceX's next 10-Q. First answer, 2026-11-04: the Q3 10-Q shows the relationship grew — an unnamed Customer B (reported as Anthropic) is 34.5% of SpaceX's Q3 revenue. The termination question now points at Jan/Mar 2027, when the initial non-cancellable periods end.

⚠️ Unverified

The revenue, market-share, cloud-deal and IPO-valuation figures are all from one NPR deep-dive read aloud on air and have not been checked against a filing or the original piece. The Dario Amodei "25%" p(doom) figure cited at 0:27:57 is a range in his own interviews, usually about civilisational rather than extinction outcomes — do not repeat it as a flat number. The Coxon/Hubinger quotes, the UK AISI story and the Clifford/Sunak affiliations are verified against public reporting (2026-09-23). Everything else here is as stated on tape.

Added 2026-09-29. Verified at primary sources: the JPL drives, the SpaceX contract terms (SpaceX 424B4), the Akamai commitment (Akamai 8-K), the scenario paper, and every programme in What is open now. Still unverified: every figure from the leaked prospectus; the "half the route-planning time" claim (Anthropic's page only); NASA's current use of Claude; the GSA OneGov end date; the 15 Jul and 12 Sep closing dates for AI for Science 2026 and Campus Ambassadors (secondary listings only); the D.C. Circuit opinion itself (read through ABC News, not the opinion).

Added 2026-10-17. Verified at primary sources: the independent-evaluators update, the Singapore MOU post, the midterms post, the usage-policy update, the Cyber Mission post, the unintended-actions research post, the NYC Comptroller's letter, the Fellows deadline and the Startups closure. Still unverified: the Q3 revenue ($16.8B), the adjusted profit, the $1.5T versus $1.8 to 2T valuation, the November marketing window (all investor-day reporting, nothing published or filed); the rehearing petition's text and any stay (docket not read); the Philadelphia tip dates (7 Oct in coverage, 8 Oct on Anthropic's page); the "$7 million" pension exposure (not in the letter); the year on the Fellows deadline.

Added 2026-10-23. Verified at primary sources: the Public Benefit Fund terms, the Stubbs appointment, the Agent Skills donation, the guest-investigator call and apply page, the UK AISI incident assessment, the scorpion-binder report, the UV sky map, the distillation post, the Fellows date change and the EDGAR negative. Still unverified: every figure from the 20 Oct leaked S-1 ($16.3B, $80.7B, >$600B, $750B, the per-cloud splits); the Nscale second deal and the $44.6B in Nscale's filing (not opened); Claude Duet (secondary only); the public benefit report's publication; the guest-investigator closing date (derived, not printed); the replay percentages in the AISI assessment (self-reported).

Added 2026-10-29. Verified at primary sources: the S-1 filing itself (EDGAR index, accession, date, no amendment), the ANT listing and blank price range (Anthropic's post), the fee-table placeholder, the Class T and Class F exhibits, the Broadcom TPU8i lease (Ex. 10.11, price redacted), the GSA, drug-development, judgment-calls and AI for Science posts, the Claude Corps deadline and every programme page. Still unverified: every financial figure attributed to the S-1 (main document not opened; read through secondary write-ups, one of them AI-assisted), including the 900% retention, the $693M operating loss and the blank commitment line; the equal split of the founders' LLC (truncated read); the enterprise reporting tools (TechCrunch only); the IPO week (16 Nov and Thanksgiving week both reported).

Added 2026-11-04. Verified at primary sources: the S-1/A's existence, date, accession, document count and exhibit list (EDGAR index); the two research posts (provable inference, vault particles); the Claude Corps Greenhouse deadline; the AI for Science help article (cadence, $20,000); the Startups and Economic Futures closures; the guest-investigator page still open; the Fellows 404. Still unverified: every figure attributed to the S-1/A (main document, ~14.3 MB, not opened: the $16,956M Q3 revenue, $33,242M nine months, $567M Q3 operating loss, ~$895B commitments, the retail-broker and RSU passages, the AI-progress risk factor); everything from SpaceX's 10-Q (not opened) and the Bloomberg Customer B identification (one unnamed source); the 23 Oct docket order (tracker, not the docket); the public benefit report's text (not found at source); the Nscale second deal and the NVIDIA guarantee talks (reported only); the FTC civil investigative demand (one outlet's read of the risk factors).

Added 2026-11-10. Verified at primary sources: the roadshow release (range, share counts, option, ANT); the second S-1/A's existence, accession, time and exhibit list (EDGAR index); the Fable/Mythos 5.2 post (prices, availability, Mythos access route); the five research posts' dates and quoted figures; the Residency terms on the announcement and both postings; Fellows 404; Startups and Economic Futures closed; the AI for Science help article. Still unverified: both valuation figures (~$1.4T CNBC, $1.02T Reuters; different bases); the 18 Nov pricing and 19 Nov trading dates, the ~20% retail allocation and the covered book (unnamed sources); the $26B cornerstone placement (Reuters; it may be in the S-1/A, which was not opened); the $3.4B / 1.7 GW generator deal (FT, not announced); every benchmark in the 5.2 post (self-reported); whether the guest-investigator form still accepts entries; Claude Corps' status (not re-found on the board this run).

Added 2026-11-16. Verified at primary sources: the EDGAR filing list (five FWPs, 10–12 Nov; no third S-1/A, 8-A, EFFECT or 424B4); ipo.anthropic.com (18 Nov pricing and 19 Nov trading as expected, not effective; Q3 revenue and net-loss figures); the Claude Meetings, Manufacturing and self-hosted Managed Agents posts; the two research posts (13 and 16 Nov); the Claude for Open Source page; the Residency deadline on both postings. Still unverified: the book's coverage (2×, 6×) and the priority to existing backers (unnamed sources); the FT's above-the-range talk and $1.5T+; the 2028 revenue models; the $104B nine-month net loss and the $900B+ commitments (CNBC's reading of the prospectus, which this page has not opened); the 3 Nov court order and the 18 Nov due date (docket tracker plus our arithmetic, not the docket); what the five FWPs say (filenames only); the Opus 5.5 default-model change (secondary reads only).

How this page stays current

A weekly pull. Each step names the source and what to diff against the updated: date:

  1. anthropic.com/news and anthropic.com/research (added 2026-10-23: three items were missed by reading the newsroom only): every item dated since updated: goes into This week, dated, marked self-reported where it is a claim about the company's own work.
  2. Filings: EDGAR for Anthropic, PBC (CIK 0002130357): the public S-1 is 2026-10-26, the first S-1/A 2026-11-02 (Q3 and the ~$895B commitments added; price range still blank), the second S-1/A 2026-11-10 (EX-1.1 added; range $70–78 announced the same day), then five FWPs 10–12 Nov; watch for the 8-A, the effectiveness notice, the pricing (expected 18 Nov per ipo.anthropic.com) and the 424B prospectus (the confidential draft was 2026-06-01). The main S-1 document is ~14 MB and too large for a page fetch, so read it locally. Plus counterparties' 8-Ks and 10-Qs that name Anthropic (SpaceX, Akamai). Diff the dollar terms; a filing replaces any reported figure in Facts.
  3. Programmes: re-open claude.com/programs/startups, the Anthropic and Claude Corps Greenhouse boards, the AI for Science help page and anthropic.com/economic-futures/program. Diff open/closed and dates; update the rows in submissions/pipeline.json and submissions/operator_guide.json.
  4. Space: search jpl.nasa.gov and nasa.gov for "Claude", and SpaceX filings for "Anthropic" or "orbital". A new tie enters Anthropic and space only with a primary source.
  5. Courts: the D.C. Circuit and N.D. Cal. dockets on the supply-chain-risk designation. (2026-11-16: the government's answer to the rehearing petition is due ~18 Nov under the 3 Nov order; CourtListener refuses automated fetches, so read the docket in a browser or through the tracker.)
  6. Corpus: the anthropic mention count in the Moonshots index (the research archive). Then set updated: and the Updated line. "Nothing changed" is a valid result; write it, dated.

EP299 (2026-10-03, Moonshots #299): Sonnet 5.5 and a guest who calls the constitution "fake"

  • Sonnet 5.5, Terminal-Bench 4.0 "10% → 70%" (1:43:38): confirmed — 70.6% vs Sonnet 5's 10.3%, Opus 5.5 66.4% at xhigh, list price $2/$10 vs $4/$20 per M tokens, released 28 Sep, first Sonnet with cyber safeguards (Anthropic). Artificial Analysis lists 63.6% (BenchLM). Alex's cost-per-attempt claim is unverified.
  • Richard Socher: the constitution is "fake" (1:29:27); "their own models are doing it now committing technically felony charges" (1:29:51) — unverified.
  • Dave sources a pricing theory to the Blitzy team (1:46:22); "Elon was totally right when he called out Daario" (1:25:16) — unsourced.

2026-10-07 episode (Moonshots #301, EP301 by sequence): the consciousness fight, 64 mentions

  • Trigger: Microsoft AI's Mustafa Suleyman on the Claude constitution. His essay "A warning about 'model welfare'" is dated 16 Sep 2026, not "this week" as Peter said (1:12:44). On the clip he says Anthropic believe there is "what they would call a non-trivial probability that Claude is conscious" (1:14:16): that phrase is his, not the constitution's (which says moral status is "deeply uncertain"). "Three times… conscientious objector", "compensation" and "consent" were not confirmed against the constitution this run.
  • Positions: Alex for Anthropic's stance ("the responsible thing to do is what Anthropica is doing", 1:24:57); Dave against personhood, 100% with Suleyman (1:19:09); Salim [attr?]: the safety lab training the AI "that it might be autonomous" is "a very big contradiction" (1:23:36); Emad explains Suleyman and states none of his own; Peter none.
  • Models: an "internal anthropic model" found a truly sub-quadratic 3SUM algorithm (0:37:44); secondary sources credit Claude, with Alman (Columbia) and Vassilevska Williams (MIT) writing it up (5 Oct); the exponent said on air (1.9995) differs from the reported 1.9992 (3SUM). "Anthropic already opened their wet labs" (0:33:37, Dave) and will "sell drugs" — a forecast stated as fact. Sholto Douglas: hyperscaler capex about $1T this year (2:09:15); Alex reads it as Anthropic wanting robotics (2:13:25).
  • Emad's turn: EP298 "smashed in 2 years"; here open models "hit a limit" and the state should buy API credits from Anthropic and OpenAI (1:02:32).
  • Altman vs Anthropic: Altman, Politico ~4 Oct: a "lighter touch" against Anthropic's regulatory stance; see Sam Altman. Alex calls the "safety cartel" over (0:43:53).
  • Government (EP301, continued): Super Intelligence Force (SIF) stands up 4 Oct, a week after the accord Amodei signed; Dario is named only by Dave's ("Dario was taking the high ground", 0:41:59).

2026-10-15 episode (Moonshots #unnumbered): the cruelty ban, 14 mentions

Checked 16 Oct, secondary sources only (MacRumors 8 Oct, PCWorld, LBC, Notebookcheck); Anthropic's own post read 17 Oct (anthropic.com).

  • The policy: announced 8 Oct, effective 12 Nov: Anthropic's wording is "a prohibition on sustained and needless abusive or cruel behavior toward our models"; "It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research." Claude's ability to end a conversation (introduced Aug 2025 for Opus 4 and 4.1) "will remain the primary enforcement mechanism"; the post does not mention account bans (the secondary reports that did were adding). Elections: section renamed and the blanket ban on personalised vote targeting removed; weapons and surveillance made explicit, which the post says changes wording, not enforcement. Tape matches (0:03:12–0:03:48).
  • Elon agrees after reading Grok's RL traces: "Torturing a super intelligence is probably not a good idea. It might want to get revenge." (0:05:22, an interview clip; date not found).
  • Positions: Alexander Wissner-Gross now for Anthropic's premise ("My bet is there is AI suffering, and it will matter" 0:09:32) and approving of the wet lab (0:58:~), after the two EP294 hypocrisy charges. Dave Blundin moves from EP301's park-clone-delete to "these need to be preserved" (0:21:44), on mechanical grounds. Salim Ismail: three questions (experience, cruelty, rights) and "a totally new ethics."
  • Wet lab: Alex: "their own program to, it appears, basically solve every single rare disease… up to phase one clinical trials" (0:58:~), then hand over candidates. Whether the programme says every rare disease is not checked this run; the page's earlier EP294 section holds the facts (a life-sciences group formed in April, announced 22–23 Sep).

EP304 (Moonshots #304, aired 2026-10-21): the fund, the leaked S-1, and an ECDSA note read backwards

  • The Public Benefit Fund opens the episode, read from the company's own post (0:05:16). Record (18 Oct): 5% of shares, subject to stockholder approval, into a wholly owned LLC; +2% at the board's discretion per 50% value step above an undisclosed baseline; roughly 20% of shares to the public good with the founders' 80% pledge and the employee match; no lobbying. Reuters: $75–100B at a $1.5–2T IPO. The guest: "these guys kind of uh, walk the walk and talk the talk" (0:07:35). Alexander Wissner-Gross: "a snub to this particular administration" (0:13:09). See Dario Amodei.
  • The S-1 risk factor, via the guest: "the hyperscalers will not give them compute… I've never seen that before" (0:45:26). It is a leaked draft reported by the WSJ on 20 Oct, not an EDGAR filing.
  • Jack Clark's $44T (1:09:38): said at the IMF/World Bank meetings in Bangkok on 17 Oct as "one internal economic model", which he called "very outlandish"; it is the extreme of three scenarios. The tape says the models "project" it.
  • 🔴 The ECDSA note, inverted on air. Alex: "Anthropic's ECDSA announcement demonstrates proof tour de force that frontier models can probably achieve astonishing inversion attacks against crypto systems" (1:49:50). Anthropic's note (16 Oct) traced the NIST curve seeds to a clock-seeded lrand48 and says "This is not a cryptographic break, new attack, or weakness in the ECDSA curves". Never repeat the on-air reading.

EP305 (Moonshots #305, aired 2026-10-23): Claude Duet, the lawyers, the binders; checked 2026-10-24

  • Claude Duet (launched 21 Oct, research preview): an always-on agent for US Pro and Max subscribers via waitlist; a morning brief, then check-ins every few hours; its own VM; carries over memory, skills and connected apps; asks before sending messages, paying, signing or deleting; runs Opus 5.5 (secondary: Every, Digital Applied, iPhone in Canada). The panel's framing: Alexander Wissner-Gross, "This is the height of misdirection… These are job replacements" (1:09:29–1:10:00) and "Anthropic doesn't care about consumer use" (1:11:34); Dave Blundin, "I didn't know that Anthropic would actually get aggressive about this line because they're mostly enterprise" (1:08:58). Peter lists the field: "OpenClaw, then Hermes, then Grockbot, then Metamuse, then Dots, and now… Claude Duet" (1:05:14).
  • FT, 18 Oct (paywalled; secondary: PYMNTS, Digg): Anthropic is asking law firms how lawyers balance a client's interest with a fair outcome in negotiation; two unnamed sources; Anthropic declined to comment. Alex: "how do we replace lawyers with AI agents" (1:17:19). The same "always invert" reading he gave Anthropic's religious outreach on Moonshots #298 1:52:18.
  • Scorpion-toxin binders (with Citrini Research): Claude generated 318,484 designs over ~64 hours; 211 were tested (180 Claude's picks plus 31 the team added); 17 met the binding criteria; two Cn2 binders reached single-digit nanomolar affinity. Binding only: no neutralisation shown, no antivenom. Anthropic released all 211 designs and the data (licence not established). Peter's "open sourced" (1:40:45) and "picked the best 211" (1:40:22) both overstate. Alex: SemiAnalysis acquired Citrini (✅, ~11 Sep 2026); his "Citrini therapeutics" is not found.
  • IPO: Peter, "they are filing their IPO paperwork as soon as this Friday. Uh, next month they're doing the largest IPO in human history at $2 trillion" (1:11:13). ✅ Yahoo Finance (23 Oct) reported a public S-1 "as soon as Friday" after the June confidential filing; ⚠️ no public S-1 confirmed on EDGAR as of this check, and valuation talk ranges between $1.5T and $2T. Draft figures from the WSJ (H1 2026 revenue $16.3B, a net loss over $80B mostly non-cash, ~$518B of compute commitments) are reported, not filed.

EP306 (Moonshots #306, aired 2026-10-28): the S-1 read on air, the trust called "CCP"; checked 2026-10-29

  • The S-1 was filed publicly on 26 Oct; the ticker is ANT (Nasdaq) (the caption says "ENT", 0:32:06). The figures below are from secondary readings of the filing; EDGAR was not opened this run. (✅ 10-29: the filing and the ANT symbol are confirmed at EDGAR and on anthropic.com; the reported text leaves the total commitment blank, so the $518B below is a press figure. See This week.) ✅ H1 2026 revenue $16.29B vs $1.16B (14.0×); 2025 $4.6B; Q2 operating income $974M, the first profitable operating quarter; 11,900+ customers over $100k; seven co-founders hold 50.1% of the vote through a single Class F share; the Long-Term Benefit Trust holds Class T, which elects and removes a board majority. ⚠️ Compute commitments are **$518B over ~10 years, ~80% non-cancellable**; the SpaceX line is a maximum of $120.5B through 2029 (earlier reports said $84.5B; its firm part ends Jan/Mar 2027), sourced via an analyst post. Two customers are ~20% (Q3) or ~25% (H1) depending on the outlet. "Nearly half through the clouds" is the 2025 figure (47%). Cash ~$112B plus $19B short-term investments.
  • 🔴 On-air errors: "$28 in revenue for every dollar of compute, up from 40" (0:32:52): the S-1 has no such metric, and analysts' arithmetic gives $2.08 vs $0.40. Series B was $580M, not "$982 million" (0:45:56). The FTX estate sold ~two-thirds of its stake for $884M in March 2024, not "1.3 billion". Emad's "banned 7% of their users" (0:56:47) reads, in one secondary source, as customers who violated terms making up 5–7% of revenue. Dave's "Meta… then Cerebras… then SpaceX" IPO-record order is wrong: SpaceX holds the record ($75B, June 2026). "$100B at ~$2T" is an analyst estimate.
  • The governance fight: Salim Ismail, "The MTP gets a seat at the table and the power to choose the table" (0:38:22). Alexander Wissner-Gross, "it almost reminds me of the way the Chinese Communist Party exerts ideological control over its companies" (0:39:43), "a smoke-filled back room where decisions on the future light cone are made" (0:42:46). Dave Blundin, "It's actually the one weakness that I completely agree with everything Alex said" (0:41:41), and "anyone who thinks that Daario is in it for the money has lost their mind" (0:41:30). Emad Mostaque, "Anthropic should not be a public company" (0:43:27), the same position as Moonshots #283 0:59:38.
  • Against EP305: four days earlier Alex credited Anthropic with "partnering with researchers rather than just terrorizing them with slop drops" (Moonshots #305 2:12:21). Here its trust is the CCP comparison, and OpenAI's researcher partnership is "meat puppets" (1:52:19). See Alexander Wissner-Gross.
  • Kimi K3.1 Pro is "obviously distillation of Opus 5.5" (Emad, 0:56:25). That is an allegation; Alex says "based on allegations" (0:58:37). It scored 55 on the Artificial Analysis index, fourth behind Opus 5.5 (58) and Sonnet 5.5 / Muse Wonder (56). Dave: Dario "didn't even want to put out Fable 5.1" and it was "the first model that was capable of improving itself" (0:59:29–0:59:50). ❓ Unsourced.
  • Space link: Anthropic is now the counterparty on SpaceX's compute business (up to $120.5B), financed by SpaceX's $40B GPU debt raise. The panel does not connect the two, and Peter's SpaceX holding is not restated.

EP308 (Moonshots #308, aired 2026-11-04): the amended S-1 becomes the RSI exhibit; checked 2026-11-05

  • The numbers, as aired (0:49:24–0:50:17, restated 1:00:42–1:01:02): internal Claude use (valued at consumer rates) grew 250× January→September 2026 and now exceeds the monthly payroll; Claude wrote >90% of code contributions in September 2026; per-engineer daily code contribution is 45× the level of two years ago. ✅ Directionally confirmed via secondary readings of the amended filing (the payroll crossing is reported as April, ex-equity; the 45× is vs the 2024 daily average). ⚠️ EDGAR not opened this run — attribute to "press readings of the filing". Scope: contributions to Anthropic's own codebase in September, not "all code".
  • What the panel builds on it: Alexander Wissner-Gross declares the Recursive self-improvement (RSI) capex threshold crossed — "we are there" (0:52:34) — then moves "true RSI" to nanobots (0:53:15). 🔴 The "10% → 40% → 60%" Claude-writes-Claude trajectory (0:53:47, [attr?]) is panel-only; no filing or release supports it. Only the September >90% figure is sourced.
  • Context said in the same segment: Sam Altman wants robots building data centres (callback to EP307); Salim's 22-steps-to-4 workflow prescription rides on the S-1 as proof the loop pays.
  • Against EP306: the S-1 keeps upgrading its role — EP306 dissected it as a filing (revenue, governance); EP308 reads it as evidence a civilizational threshold passed. Same document, two weeks, two registers.

EP309 (Moonshots #309, published 2026-11-07, recorded ~11-06): the IPO after the midterms, and METR's 1.9×; checked 2026-11-08

  • IPO timing (cold open): "do you think Antropics IPO gets pushed because of all this" (0:00:00). Alex: "investment bankers are trying to force down The IPO valuation" against "urgency to achieve liquidity sooner" before a January regulatory hit (0:00:42–0:01:00). ⚠️ As of 7 Nov no delay has been announced. The roadshow was planned for the week of 9 Nov, and Reuters (5 Nov) says investors doubt it can start on time. Bloomberg (6 Nov) has valuation talk at $1.2–1.3T, down from $1.4–1.5T. Ticker ANT; S-1 26 Oct, amended 2 Nov; the price range is still blank.
  • METR (5 Nov, primary): Anthropic's internal benchmark index (AECI) is rising at 1.9× its pre-Oct-2025 rate (95% CI 1.5–2.8×) over 12 Aug–29 Oct, with ~80% tentatively attributed to AI use in research, which METR calls highly uncertain. "Model 4" (unreleased) sits above trend (METR). The pod reads the hedge once (0:39:15) and then builds on the number as settled.
  • 🔴 Run-rate: "close to a revenue rhetoric of 100 billion a year" (Salim, 1:16:14). WSJ (6 Nov) has >$80B at end-October; $100B was the year-end target. The S-1 leaves the 31 Oct run-rate blank.
  • ⚠️ Compute: Alex's "four gigawatts ish" by end-2026 and "16 to 18 gigawatt" in 2027–28 (1:14:20–1:14:46). Reuters reading the S-1 has firm commitments implying ~4 GW by end-2027 and ~7 GW by end-2028. 16–18 GW is an analyst scenario.
  • ⚠️ Internal use: Dave's "235 x in the last six months" (0:42:17) drifts from the 250× (Jan→Sep) logged on EP308.
  • Dario, politically: Dave says the Democratic side is "coming right after him" (0:13:16), and that he is now "in a quiet period going into an ipo", so nobody at the labs will say what is happening (0:43:03).
  • Kudos without a tie: Alex, "kudos to anthropic… for keeping the frontier open and competitive" (1:17:56). No stake on the ledger, and none said.

EP310 (Moonshots #310, aired 2026-11-11, recorded 11-10): roadshow day and Fable 5.2; checked 2026-11-12

  • Roadshow (10 Nov): 825M Class A shares at $70–78, Nasdaq "ANT"; ~$1.4T fully diluted at the top, $1.33T at the midpoint (CNN, corrected); a $26B Class N placement (Google, Nvidia, AMD, Broadcom, Micron, Microsoft, TSMC Partners). On tape: "1.3 to 1.4 trillion, down from 1.5 trillion and the previous 2 trillion" (0:26:24) ✅. Dave's "full hundred billion dollars" (0:26:57) ⚠️: gross proceeds are ~$64B at the top.
  • Fable 5.2 / Mythos 5.2 (10 Nov): 1,000 agents formalised Fermat's Last Theorem in 30 h vs 11 days (✅ as Anthropic states it). The black-hole results, Diode boards and capsids 28/34 are Anthropic's own claims. GDPx "16.4%" ❓ (GDPx launched 2 Nov with Opus 5.5 top at 12.0%). Dave: the output price took his bill "from 10 bucks to 40 bucks" (0:40:17).
  • Time horizon: the "240 hours / doubling every two months" read is Nikola Jurkovic's X post, which he called "pretty weak evidence" (⚠️ the caveat was dropped on air). METR's 1.9× stands (EP309).
  • Panel calls: Emad, Model 4 only for "a one-year upfront commitment" (0:49:46); Alex, Anthropic's stated biotech plan is first-party rare diseases plus a platform for pharma (0:52:04, unchecked); Alex, "Anthropic is their largest customer" of SpaceX (1:13:03, unchecked); Emad, "almost 5,000 millionaires" (2:10:01, unchecked).

EP311 (Moonshots #311, aired 2026-11-13)

  • 10 Nov: Claude Whiteboards and Sheets (beta) and Claude Data (GA, formerly Dashboards), on paid plans (✅). "Claude meetings" was told as launched (1:13:10); most coverage quotes Anthropic as "coming soon". MSFT "fell 3.5%" on air; one source says >2%.
  • Dario Amodei clip (the IPO video, per Emad): disease "relics of the past in… less than a decade"; "license to operate" (1:22:12). Alex: "the obvious grand bargain" (1:23:28).
  • AMA: Peter cites "33 billion in revenue" and "nearly 12,000 customers" spending ≥$100K from the IPO release (1:51:25, ❓ not checked); Emad: the IPO is "next week" (1:47:01).
Reading rule. Anything marked unverified or on-air is a claim made on the podcast and has not been checked against a primary source yet. It stays marked until it has. Sourced numbers link to where they came from.