Source: Matthew Berman, “12 months to the singularity” (raw/12_months_to_the_singularity.md, https://www.youtube.com/watch?v=lVbQqjAJpkE; sponsored by JetBrains), reading METR’s report on screen · The Neuron Daily, November 6 (raw/newsletter-theneurondaily-com-2cee76fc49.md) and November 4 (raw/newsletter-theneurondaily-com-5403db9871.md) · The AI Daily Brief, “AI Lurches Towards RSI” (raw/AI_Lurches_Towards_RSI.md, https://www.youtube.com/watch?v=Es59icF3nZg) · AI Explained, “We may release models we can’t control” (raw/We_may_release_models_we_can_t_control_-_Amodei_Altman_..._and_Gemini_Carbon.md, https://www.youtube.com/watch?v=0Kw6_DTZPrY). Also The AI Daily Brief, “Companies Are Spending 5x More on AI Now” (raw/Companies_Are_Spending_5x_More_on_AI_Now.md) and Moonshots EP309 (raw/Anthropic_s_IPO_in_Question_Midterms_Hit_AI_Nasdaq_Drops_5%_and_RSI_Accelerates_309.md), added from another writer’s proposal. METR’s report was not fetched; quotes are hosts reading it aloud.
METR, the independent AI evaluation group, published “Independent assessment of AI R&D acceleration at Anthropic, October 2026” in early November. Most of the METR staff on the project now think Anthropic is improving its models on well-specified tasks at about twice the rate of March 2024 to October 2025, that roughly 80% of that acceleration comes from AI itself, and that nothing they saw will stop it compounding over the next 12 months. Anthropic controlled what METR could see and what it could publish, and cut a planned check of its compute allocation midway. The same week, Anthropic’s amended IPO filing said it may soon cross its own recursive-self-improvement (RSI) danger threshold.
Key Takeaways
- About 2x (1.9x), with a wide band. Moonshots gives the figure as 1.9x from an 11-week study. METR’s staff “would not be surprised if the current rate of growth was closer to 1.3x or 3x the historical rate.” The baseline is March 2024 to October 2025. The Neuron’s one-line version: capabilities “improving about twice as fast as a year ago.”
- Mostly caused by AI, METR thinks. The acceleration that began in late 2025 was “likely due to non-AI factors” (more compute, more hiring, human breakthroughs). By October 2026, “we guess that the majority, perhaps about 80% of the current acceleration is due to AI.”
- No bottleneck in sight for a year. “We did not see evidence for significant bottlenecks in algorithmic progress, training compute, experiment compute, or data that will take effect in the next 12 months.” Anthropic redacted METR’s analysis of several specific bottlenecks.
- Benchmarks are saturating. Anthropic “appears to have made as much progress on benchmarks in the past year as they did between October 2023 and October 2025,” and has had to make its internal benchmarks harder.
- The access was limited. 2.5 full-time-equivalent METR staff from mid-August. Anthropic could approve or reject every data request, and Berman says it decided what was published. Berman reads METR as saying this was not the employee-like-access program in Anthropic’s October 15 update on independent evaluators. ^[ambiguous: The AI Daily Brief introduces it as that program’s first report]
- Anthropic’s own filing points the same way. New language in the S-1A says internal estimates indicate Anthropic “may imminently meet the RSI danger threshold”: as much progress in one year as two years of the 2018-2024 baseline (Andrew Curran, via The AI Daily Brief).
What METR was asked
Three questions, per Berman’s reading:
- Has the growth rate of Anthropic’s internal capabilities accelerated?
- What share of any acceleration is attributable to AI?
- Are there important bottlenecks that could slow growth or limit acceleration in the near term?
Evidence used: interviews, self-reported uplift, headcount, lines of code, experiment usage and “qualitative reflections from leadership.” METR also says it received “convincing evidence that the acceleration was not due to an acceleration in the scale of training compute”; Anthropic redacted the details.
Findings
- The capability curve bent at Opus 4.5. Anthropic’s internal capability index (AECI, its fork of Epoch’s ECI) is roughly linear through late 2025, bends at Opus 4.5 and keeps steepening through Mythos. Berman also shows an explanatory chart that METR labels illustrative, not measured data.
- Not a one-off jump. From the appendices: Anthropic’s newer, unreleased internal models “make a permanent slope change look more plausible than a one-time jump at Mythos preview.”
- Autonomous research is close. AI “is not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process,” but researchers described “significant signs of life,” and say models are becoming useful for autonomous experimentation for the first time.
- Compounding risk. “We believe there is a significant risk that further capabilities progress could lead to even faster acceleration over the coming year.”
- Staff quotes (anonymised): “could definitely endorse a greater than 2x speed up”; “Now only bottlenecked on ideas and compute. No longer bottlenecked on code”; “greater than 2 years of progress in one year and I don’t think four is crazy.”
More detail from other relays (added 2026-11-09)
Sources: The AI Daily Brief, “Companies Are Spending 5x More on AI Now” (raw/Companies_Are_Spending_5x_More_on_AI_Now.md, https://www.youtube.com/watch?v=INC89yHLexM); Moonshots EP309 (raw/Anthropic_s_IPO_in_Question_Midterms_Hit_AI_Nasdaq_Drops_5%_and_RSI_Accelerates_309.md, https://www.youtube.com/watch?v=WF-ADiC7Khs).
- The headline figure. Moonshots gives it as “approximately 1.9 times its historical rate,” from an 11-week METR study, with “roughly 80%” of the acceleration from AI assisting AI research and no bottleneck seen over the next year. Eleven weeks fits Berman’s mid-August start.
- Unreleased models. METR saw analysis of three unreleased models trained since July; the most recent was completed in September or October and “continued the trend of accelerated improvement” (The AI Daily Brief). On Moonshots, Alex Wissner-Gross points to a model labelled “model 4” on METR’s charts sitting above the fitted exponential trend, and to commentary that progress is within 10-15% of the AI 2027 scenario’s prediction for November 2026. Those are his and others’ readings of the charts.
- How METR describes the acceleration: “we were told that R&D is accelerated primarily via AI assistance rather than fully autonomous contributions, but there are some signs that fully autonomous contributions will play a larger role soon.” It warns that further acceleration “could lead to a sudden emergence of extreme and risky capabilities” (The AI Daily Brief).
- Access, two framings. The AI Daily Brief says the access METR received “surpassed their expectations” and introduces the report as the first from the evaluator Anthropic onboarded “last month” with employee-level access. Berman reads the report itself as saying this project was not that program and did not involve full employee-like access. Both may hold if the program announcement and this project overlapped in time but differed in scope.
- Anthropic’s response. Co-founder and responsible scaling officer Sam McCandlish joined X to discuss it: “Meter’s work here provides the public with an independent external view into how this is happening inside Anthropic.” He acknowledged criticism that Anthropic redacted some of the most important evidence and said Anthropic will publish a supplemental report on its own analysis in the coming weeks.
- Reactions (The AI Daily Brief):
- Ryan Greenblatt (METR): “a big update towards ASI within one to two years.”
- Nathan Calvin (general counsel, Encode AI): “extremely unnerving update.”
- New York assemblyman Alex Bores (captioned “Boras”): “Congress should ban RSI this year if possible.”
- Nathan Lambert, sceptical of the method: “a lack of good enough benchmarks is a bottleneck, at least for the type of RSI described as an intelligence explosion.”
- Critics restated that METR is close to Anthropic in ideology, and NLW notes the report does not address sandboxing or the cybersecurity concerns that prompted outside evaluation.
Added 2026-11-11: dates, the redaction, and Google’s version
Sources: The Artificial Intelligence Show ep. 250 (raw/Ep._250_-_Midterm_Voters_Reject_Data_Centers_AI_Speeds_Up_Anthropic_s_Research_Nvidia_s_Spy_Probe.md, https://www.youtube.com/watch?v=2XY3q56sD04), with Mike Kaput reading METR’s post and Paul Roetzer reading the X posts; The Epoch Brief, November 11 (raw/newsletter-epoch-ai-fa4c3ebf04.md).
- Dates. METR’s post: “from August 12th to October 29th, 2026, METR staff performed a high access evaluation at Anthropic to assess the degree to which AI is accelerating AI R & D.” That is about eleven weeks, matching Moonshots’ “11-week” figure.
- How much was redacted. METR on X: “We submitted a 100-page full report of our assessment to Anthropic. For the public version, Anthropic redacted a majority of the evidence we found most informative. As a consequence, the public must rely on our judgment more than we consider ideal.” Anthropic said the redactions were “aimed at removing confidential information that could accelerate other AI developers.”
- McCandlish’s post, more fully. “Many of METR’s conclusions match my own. AI is speeding up our research, but it’s hard to say exactly how much.” On the redactions: “sharing exactly what is driving the speed up isn’t the right thing to do, both because it’s commercially sensitive and because it risks handing out a roadmap to others to go faster.” Anthropic’s promised report will cover “the speed of AI progress within Anthropic, including where we stand against our responsible scaling policy”, plus new safety and security steps. His role is given differently: “responsible scaling officer” above, while Roetzer describes him as former CTO and now chief architect, and his own post says he leads “much of our AI safety work”.
- A scale for “two years in one”. Kaput: METR describes the past year’s progress as roughly the distance from Claude 2.1 to Sonnet 4.5, achieved in half the time.
- Google describes the same thing. Sergey Brin, at an October 15 Business Insider roundtable on Gemini 4 Argon, said models now help build their successors and Google was “starting to get onto the RSI treadmill”, citing the Flash models released before Gemini 4. A Gemini RSI lead posted that post-training demand had shifted “from asking for raw chips to asking for Argon quota”. DeepMind’s Pushmeet Kohli: “a genuine phase transition in the pace of scientific discovery”, with releases due within weeks. (All relayed by Roetzer.)
- An outside data point. Epoch’s Capabilities Index gave Claude Fable 5.2 a record 174 (previous high GPT-6 Astra, 169), above the trend since reasoning models arrived in late 2024. Epoch: “There has been a lot of recent discussion about AI contributions to AI R&D … but Fable’s score is just one data point.” See Epoch AI, November 2026.
- Roetzer’s reading (opinion): progress is running ahead of even the most aggressive AI 2027 projections, and business leaders should treat that as a reason for urgency.
What METR could not check
- Compute allocation was cut from scope. The contract originally included verifying how Anthropic splits compute between capabilities and safety. Anthropic “changed the scope of this project midway”; per METR, Anthropic said it wanted “to keep the engagement focused on risk assessment for the time being” and would revisit compute verification later. METR reached no conclusions on compute use.
- Negotiated data. “In practice, we often negotiated with Anthropic staff and compromised towards data that was informative, not extremely sensitive and easy to share.”
- Berman’s speculation, not METR’s: the scope change may relate to the IPO.
How this fits earlier readings
- This is a change over time, not a contradiction. Anthropic’s Mythos 5 system card said no sustained, AI-attributable 2x acceleration had occurred, and METR’s external testing then found Mythos 5 unable to fully automate multi-week R&D (see Fable 5 and Mythos 5). METR now puts the acceleration at about 2x. Anthropic’s own AECI slope ratio after Mythos Preview was 1.86x to 4.3x depending on the breakpoint (Mythos Preview system card).
- Two different METR engagements. Anthropic’s Position records a six-person METR team embedded with employee-like access. This report says it was a separate 2.5-FTE project without that access. ^[inferred: the two descriptions refer to different engagements]
Same-week RSI signals
- Anthropic’s S-1A (amended Monday, November 2). Beyond the threshold language above, Anthropic says it is considering changes to its Responsible Scaling Policy because it does “not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding” (The AI Daily Brief). A risk factor quoted by AI Explained: “if we become less able to detect or assess misalignment as our models advance, we may deploy models whose behavior we cannot adequately verify.”
- The “26% of research” chart reached Wall Street. The Wall Street Journal ran Anthropic’s chart showing AI leading 26% of research, which Anthropic first published in September from August data (The AI Daily Brief). That figure’s baseline is disputed: see the contradiction callout in Anthropic on Recursive Self-Improvement.
- Google says the same. Sundar Pichai on Alphabet’s earnings call: “We’re at the point where AI is helping us increase our training and launch velocity. I would expect the next 12 months to see continued acceleration because of RSI.”
- Anthropic staff expectations (a survey quoted by AI Explained): 91% think AI could make a Nobel-level discovery by 2030; on average they expect AI to do 75% of their own work by then, and 95% of the work of training a frontier model by end of 2030, with some human role remaining between the 10th and 90th percentiles. Co-founder Jack Clark has moved his estimate for full RSI being possible to 2027, from 2028.
- Chip design. Anthropic told IPO investors Claude helps its 15-person chip team design custom silicon faster, based on its own estimates (The Information, via The Neuron).
- Dwarkesh Patel builds a timeline on it (added 2026-11-13). His video essay feeds this report’s 2x figure into a takeoff model and gets 2027 as the naive date for AI matching the best human AI researchers; his own median is 2028. He also relays an OpenAI September post: the heaviest 10% of OpenAI researchers went from about 7,000 a day between early summer and end of August. See Recursive Self-Improvement Has Started.
- A sceptic on what follows. NLW (The AI Daily Brief) doubts that faster models translate straight into revenue: enterprises use “a vanishingly small portion” of available intelligence because “inertia is real” and “adoption is hard.”
Try It
- Shorten your model re-evaluation cycle. If capability is compounding at about 2x, a model routing decision made in the summer is already stale. Re-run your own task bench on each new release rather than quarterly. ^[inferred] Method: Picking the Right Model.
- Quote the scope with the number. METR’s 2x is for “well-specified tasks” inside one lab, with lab-controlled access. Don’t present it as a general rate of AI progress.
- Watch the RSP update. Anthropic said it is considering changes; any new policy will say what human oversight it keeps over AI-driven model development.
Fable 5.2 (2026-11-09): capability held back on AI R&D
- Anthropic’s launch page says that “in relevant areas, including accelerator kernel optimization and improving large language model architectures, Fable 5.2 and Claude Mythos 5.2 perform at about the level of Opus 5.5.”
- The system card (now in the vault, checked 2026-11-11): Anthropic believes internal use of Mythos 5.2 “does not accelerate our aggregate AI R&D progress by a factor of 2”, so it does not meet the RSP threshold, “though this assessment is uncertain”. It withholds CoBench, its root-cause diagnosis benchmark, “because we are updating the evaluation”. METR’s pre-deployment testing (Section 2.3.6) ran on a snapshot stronger at AI R&D than the released model, called ‘Model 3 Early’ in METR’s October report, with observations dated September 23.
- The AECI trend bent upward. A slope change fits Anthropic’s capability index better than a one-time jump; Anthropic calls this “a meaningful update toward the possibility of durable AI-driven acceleration”, and reports AECI for “a separate, unreleased internal research model that is somewhat more capable than Claude Mythos 5.2”.
- See Claude Fable 5.2 + Mythos 5.2.
Open Questions
- METR’s report was not fetched. All figures are Berman reading the screen; the report’s URL is not in the sources. Partly answered 2026-11-11: the assessment period was August 12 to October 29, 2026 (METR’s post, read by Kaput).
- The compute-efficiency figures Berman reads from the S-1 (“67x” in Q4 2025, “break even” in Q1 2026, “2x” in Q2 2026) are garbled in the captions and are not used here.
- Will METR’s compute-allocation check be reinstated? And what will Anthropic’s promised supplemental report show about the redacted evidence?
- Did this project run under the employee-level-access program or beside it? The relays disagree (see “Access, two framings”).
- What exactly is the S-1A “RSI danger threshold” and how is it measured against the 2018-2024 baseline?
Related
- When AI Builds Itself — Anthropic on Recursive Self-Improvement — Anthropic’s own RSI data and the 26% chart
- Weight-Space Programming — the same week’s paper on a model editing another model’s weights
- Lab Slowdown Warnings — “Pacing the Frontier” and the embedded-evaluator promise this project is distinct from
- InnovationEval — Epoch’s external test of AI R&D automation
- Claude Mythos Preview — the AECI slope analysis this report extends
- The Verification Frontier — why acceleration shows up first on well-specified, checkable tasks
- Anthropic’s Position — the IPO context and the earlier METR embedding
- Recursive Self-Improvement Has Started — Dwarkesh Patel’s timeline built on this report