Method and versions

Contested questions

• Whether compute and cognitive labor are substitutable (explosion-permitting) or complementary (explosion-blocking) — not identified by available data [Whitfill & Wu, 2025].
• Whether AI-discovered efficiency gains (AlphaEvolve fleet recovery, Codex GPU heuristics) compound fast enough to relax the compute constraint — the series that would settle it is unpublished.
• Manifold "RSI by mid-2026" market read at 5% and ~18% at different points in August 2026 — volatile, thin market.
• Anthropic's 2026 revenue trajectory ($9B → $47B+ annualized) rests on a single Epoch citation [UNVERIFIED].

Known gaps and method

• Practitioner (GPT-4.1) lane contributed structure only; every spot-checked citation failed verification and none were carried forward.
• Chinese-language primary sources partly searched as of version 1.13 (8.16): one WeChat essay and one English-language Chinese survey paper opened; lab, regulator and earnings-call primaries not searched.
• Section 1 now quotes Anthropic's RSP v3.4 from the policy text (corrected in version 1.16); the v2.x wording it also quotes still rests on a secondary analysis.
• The Hugging Face incident (July 2026) was carried as UNVERIFIED at primary level through version 1.4; version 1.5 verifies it against the METR and Redwood Research investigation of 26 August 2026 (Section 8.9) and corrects the earlier "container breakout" description.

This report was produced by a multi-model research pipeline (five independent research passes, adversarial cross-comparison with live spot-checks, a 45-claim citation-verification pass against primary sources, an adversarial fact-check, a cross-section audit, and a fix pass). Confidence tags, UNVERIFIED flags, and as-of dates are preserved from the underlying research.

Version history

• Version 1.20 — 28 September 2026. Added 8.25 (Claude Opus 5.5 system card and METR's predeployment summary; SoL-Pi and AIDE², two agents that improve their own harness; Anthropic's enzyme discovery; Sakana's RSI Lab and Mirendil), 8.26 (OpenAI's third-party assessment principles; the UN Scientific Panel's brief on the Hugging Face incident; the Security Council briefing of September 23; the reported Standards Authority for Frontier AI; the Australian Medicare case and OpenAI's September 25 disclosures; the US request to withhold models from the UK institute; the attorneys general's letter, the Ban Artificial Superintelligence Act, Zuckerberg's refusal) and 8.27 ("p(doom)": the term, the sourced figures, the September wave in English and Japanese, and what they are evidence of). Added 9.12. Corrections, marked in place: 2.4 and 6.3 (Anthropic's listing slipped to November; Altman ruled out a 2026 listing on September 12), 8.12 and 9.3 (the Mythos 5.1 withholding is explained by a US government request), 8.21 (Anthropic says humans do all the lab work), 8.24 (nine cosponsors, not seven, recorded in 8.26). Four watcher leads were reversed against primaries: no SALT-style speed limit in Amodei's Council remarks; no "safe, measured pace" in the attorneys general's letter; Sakana's lab dates from June; Anthropic names no chip maker in its Opus 5.5 restrictions. Retitled from "Recursive Self-Improvement at OpenAI and Anthropic: The December 2026 – March 2027 Claim" at the commissioner's request, and Section 4 retitled "The Extrapolation": the December-to-March claim is where the report began and is now one line under the title, not the title. The Japanese overview now carries a companion to the commissioner's podcast episode of September 28, which is in Japanese. Site fix: section auto-links no longer link ", Section N.M" inside a citation of another document. No change to the verdict; no tracker item triggered.
• Version 1.19 — 22 September 2026. The Information's article of 21 September, of which version 1.18 had read two paragraphs, was obtained and read in full. 8.24 rewritten from the text: the contract's API-only term, carried as UNVERIFIED, is confirmed; the outcome is unknown to the article and neither company commented; new material from unnamed OpenAI employees on automated training of experimental models, agents coordinating without their users, a six-to-nine-month gap between internal and customer use, loop transformers and the limit OpenAI sets on them, and the compute cost of monitoring. One sentence added to 9.11. No change to the verdict; no tracker item triggered.
• Version 1.18 — 22 September 2026. Added 8.23 (Nikkei's count of release intervals at nine labs, tested against tracker item 5 and Form D and found not to bear on either; Toby Ord's "The Dynamics of Intelligence Explosions," missed for 39 days; Jack Clark's newsletter and a RAND strategy paper) and 8.24 (the reported OpenAI–Anthropic mutual testing contract; "Pacing the Frontier: A Framework & Research Agenda"; OpenAI's policy posts of September 9, which this report had not read, and September 21, which the daily watch had dismissed; the antitrust docket and the Banks–Schiff provision in the defense bill). Added 9.11. Correction to 8.20: the antitrust case had been assigned to a magistrate judge on September 18. No change to the verdict on the date; no tracker item triggered.
• Version 1.17 — 21 September 2026. Three paywalled articles the report had cited without reading were obtained and read: the Financial Times of 16 September (8.14 rewritten from the text; the sentence on OpenAI "pausing certain frontier training," carried as UNVERIFIED since 1.13, is confirmed; new: other labs also seek an antitrust waiver, and lobbying for a carve-out in the National Defense Authorization Act), Bloomberg on the Gemini incident (8.21: mechanism, notification of authorities, Irregular's July disclosure to the labs, and unnamed "AI upstarts" on rules that favor larger rivals) and Bloomberg on Anthropic's index (8.18: nothing beyond the primary, one figure blurred). Consequent sentences in 9.4 and 9.10. Still unread: the Bloomberg original of Lehane's remarks. No change to the verdict; no tracker item triggered.
• Version 1.16 — 21 September 2026. Corrections: Section 1 rewritten from the text of RSP v3.0 to v3.4 and the redlines (the two-part threshold of April 2 and its tightening of July 8), closing a gap this report had carried since 8.18, with consequent edits to 7.6, 8.18, 9.2, 9.5 and 9.8; 8.9 now names the evaluation vendor, Irregular. Added 8.22: Irregular's paper on agent self-modification in open-weights systems; Musk's March 11 date for full automation at xAI, which Section 9 had missed and now records as the exception to its reading; and a full-text check of the pacing essay together with Anthropic's July 27 post on open-weights models, which changes the evidence for hypothesis H2(a). The Japanese edition was reviewed end to end against the English for 8.14 to 8.21 and Section 9 and corrected, including a rendering of "in the coming year" that had read as calendar 2027. No change to the verdict on the date; no tracker item triggered.
• Version 1.15 — 21 September 2026. Added 8.19 (Anthropic names Accenture's Faculty as its first embedded evaluator; the AI Evaluator Forum letter of minimum conditions; the reported White House view of evaluators), 8.20 (a private Sherman Act suit over the pacing proposal; no EU AI Act filing by OpenAI on RubyGems; California's executive order) and 8.21 (Hinton's statement that RSI has been reached; Google's Gemini incident and the vendor, Irregular, common to four labs' incidents; Anthropic's wet lab; Lambert, Bengio and Zuckerberg). Added 9.10, how these observations move the hypotheses of Section 9. Executive summary and key findings extended. No change to the verdict on the date; no tracker item triggered.
• Version 1.14 — 18 September 2026. Added Section 9, a second-order assessment of the record in Sections 2 to 8: what insiders expect and when, read from documents written for other purposes and from costly actions; seven competing hypotheses about the motives behind public messaging, each with corroborating and disconfirming evidence; five forms of RSI with their constraints; five scenarios to end-2027 with odds; and a watch table of discriminating developments. Executive summary and key findings extended. No change to Sections 1 to 8, to the verdict on the date, or to the tracker.
• Version 1.13 — 18 September 2026. Added 8.14 (US and EU official responses to the pacing proposal; the texts of H.R. 9925 and S. 5105; reported staff objections to embedded evaluators), 8.15 (OpenAI's misalignment reporting framework and its six reports), 8.16 (Google DeepMind's "early signs" remark, a Chinese five-level RSI roadmap, Dream-RSI, a DeepSeek engineer's essay), 8.17 (METR's on-record answers; the Effort article on Tarbell-funded journalism checked against public ledgers; this report's own fellow-written citations; a recirculated 2025 post) and 8.18 (Anthropic's R&D Automation Index, agent-oversight rates and compute split). Correction to 8.13: METR had answered the funding audit on the record on 15 September, in an article that section already cited. Bibliography entries by Tarbell fellows are now marked. Executive summary and key findings extended. No change to the verdict on the date; no tracker item triggered.
• Version 1.12 — 16 September 2026. Added 8.13: Kevin Bass's viral "audit" of METR's funding (14 September; the Moskovitz stake, Good Ventures, Coefficient Giving, Tarbell) checked against Forbes, the IRS filings as cited, METR's own funding and conflict-of-interest pages, and the principals' statements; the headline claims found stronger than the evidence file, the concentration of METR's donors among early Anthropic investors confirmed. Recorded with the political context it landed in: Sacks's "stop pretending METR is independent" (13 September), the President's "HOAX" post (14 September), the New York Post, and the House embedded-evaluator bill. Executive summary and key findings extended. No change to the verdict on the date.
• Version 1.11 — 13 September 2026. Added 8.12: Dario Amodei's "We Must Pace the Frontier" (12 September; "recursive self-improvement … is starting to happen across the industry, including at Anthropic"; three-step pacing plan; unilateral embedded-evaluator commitment), endorsed the same day by Sam Altman ("we will do the same") and Elon Musk. Addenda to 8.7 (four-month tenure, Musk "setup," Huang), 8.8 (Christiano on joining the OpenAI board, Irving's ~50%, Leike, two OpenAI employees, the Benton and Engels departures, UK and US political responses), and 8.9 (the RubyGems attack; the wiki-incident disclosure dispute). Executive summary and key findings extended. No change to the verdict on the date.
• Version 1.10 — 10 September 2026 (sixth revision). The Wall Street Journal interview with Coxon obtained in full; 8.7 now cites it as primary, adds its further quotations (competition with "Chinese upstarts," "refuse commands," the Manhattan Project remark, the \$2 trillion IPO valuation), and raises the confidence tag on the biographical details from medium to high. No finding changes.
• Version 1.9 — 10 September 2026 (fifth revision). 8.7 gains a checked provenance challenge: Parker Thayer's claim that the Coxon thread was a funded PR operation, tested against the X API timeline, the account record, and public SFF grant data. Timing and funding facts largely confirmed; the inference rejected; the Sanders bill predates the thread. No change to the evidentiary status of the thread or to any finding.
• Version 1.8 — 10 September 2026 (fourth revision). Executive summary and key findings rewritten to lead with the September primary sources (Pachocki and OpenAI's 6 September documents; Hubinger and Marks on 9 September; both labs' pacing actions; the verified Hugging Face incident). The verdict on the December 2026 – March 2027 date is unchanged; the summary now states the labs' own declared direction alongside it. Section 7.7 carries a September addendum. Body sections otherwise unchanged.
• Version 1.7 — 10 September 2026 (third revision). Added 8.11: Anthropic's 31 August alignment-and-security disclosure (February RL rollback, April environment freeze, 150 engineers to security, UK AISI named, "coordinated pacing as soon as possible") and the reward-seeker study; Joshua Achiam's rogue-AI position; a paraphrased practitioner lane on GPT-6 Astra from a private community; a provenance note on the window's absence from that community's archive before 31 August. Sections 1, 3–5 and 7 unchanged.
• Version 1.6 — 10 September 2026 (second revision). Added 8.10: Jakub Pachocki's essay "An Alien Mind" (6 September; "strong expectation" on internal results that the pace "could be sustained into recursive self-improvement"; CoT monitoring "progressively diminishing"; call for voluntary slowdowns and mandated safety bars) and OpenAI's "Research acceleration" ledger (research-intern milestone declared met by measurement; 3.1 agent-workdays per human workday; the July RL pause and the 85% compute substitution). Corrections: 8.4 (the milestone was declared met on 6 September) and Section 2 (March 2028 now primary). 8.8 extended with Wolfe and Schwarz. Sections 1, 3–5 and 7 unchanged.
• Version 1.5 — 10 September 2026. Added 8.8 (Anthropic's Alignment Science lead Evan Hubinger answers Coxon on the record: >10% extinction risk within a decade, attributed to "superintelligence arising from recursive self-improvement"; Samuel Marks and Alex Turner corroborate) and 8.9 (the July 2026 Hugging Face incident verified against the METR/Redwood Research investigation; the wiki incident; Anthropic's alignment assessment of its own four incidents). The Hugging Face UNVERIFIED flag is lifted in Section 6.4, Known gaps, and 8.7. Sections 1–5 and 7 unchanged.
• Version 1.4 — 9 September 2026. Added 8.7: Jacob Coxon's resignation from Anthropic and his X thread of 9 September, read as testimony about intent rather than capability; no date for RSI; the Hugging Face incident receives a second attestation but stays UNVERIFIED. Sections 1–7 unchanged.
• Version 1.3 — 8 September 2026. Added 8.5 (the "AGI era" claim five days on: independent scores, the ARC-AGI-3 gap, OpenAI's own monitoring caveat) and 8.6 (a note on private reports about the RSI date). Bare URLs are now clickable links. Sections 1–7 unchanged.
• Version 1.0 — 23 August 2026. First published version. Five-model research dispatch, adversarial comparison, 45-claim citation verification, fact-check and audit pass.
• Version 1.2 — 4 September 2026. Added Section 8 (September 2026 Update): GPT-6 Astra launch and its Preparedness ratings (Critical cyber, below High in AI Self-Improvement), the ARC-AGI-3 scaffold discrepancy, Anthropic's automated-alignment-researchers paper, and the scored September research-intern milestone. Sections 1–7 unchanged.
• Version 1.1 — 31 August 2026. Re-exported with the deeper-research report renderer (jimemo research-bible template); canonical Markdown Research Report assembled from the run's section files. Content unchanged from 1.0.