Takeaway: On November 5, 2026, the independent evaluation group METR published a public report from an 11-week, high-access review inside Anthropic (August 12 to October 29). METR’s authors now believe that, as of October 2026, Anthropic is likely improving its models on well-specified tasks at around twice the rate seen between March 2024 and October 2025 — with high uncertainty (they would not be surprised if the true figure were closer to 1.3× or 3×). Measured another way, Anthropic’s internal AECI benchmark index has grown at an estimated 1.9× its pre-October 2025 rate (95% confidence interval 1.5–2.8×), which METR describes as roughly two years of historical progress in the last year. The group guesses that a majority — “perhaps ≈80%” — of the current (October 2026) acceleration is due to AI use, and it did not identify clear bottlenecks that would block even faster compounding over the next 12 months. Anthropic redacted “a majority of the evidence we found most informative.” Mainstream coverage followed, including Robert Hart’s November 7 report in The Verge.
This is a measurement-and-transparency story about AI accelerating AI research. It is not a product launch recap of Claude Personal Finance, FlashProve, or Claude API message threads, and it is separate from this week’s developer-security write-up of Claude Code settings files leaking secrets on GitHub. It also sits beside, rather than repeating, OpenAI’s own cybersecurity-model news on GPT-6 Cyber and the 948 Linux kernel LPE pipeline.
What METR was invited to look at
METR (authors Tom Cunningham, Nate Rush, Parker Whitfill, and Hjalmar Wijk, with two additional staff supporting) ran the review with about 2.5 full-time-equivalent researchers. Anthropic offered what METR calls unprecedented access to internal R&D data: pretraining and reinforcement-learning metrics, training-compute figures by domain, benchmark scores for internal models, AI speedup analyses, forecasts (later redacted), and interviews with staff.
The contracted questions were straightforward:
- Has capability growth accelerated inside Anthropic?
- How much of any speedup is attributable to AI rather than compute, hiring, data, or other factors?
- Are there near-term bottlenecks — algorithms, training compute, experiment compute, or data — that would stop further acceleration in the next year?
METR stresses this engagement is not the separate October 15 program in which Anthropic promised some independent evaluators “employee-like” access. Reviewers did not get unmediated access to documents, Slack, or unreleased models, and they did not sample non-public models. Information typically arrived within a few days of a request; a small number of requests were refused as too sensitive. Anthropic provided company devices and free tokens (used mainly for data analysis and visualization). METR took no payment.
The original contract also included a check of how Anthropic splits compute between capabilities and safety. Anthropic later narrowed the scope to R&D acceleration only, before METR reached conclusions on compute allocation, saying it wanted to keep the engagement focused on risk assessment and would revisit compute verification later.
The numbers METR can share
| Finding | Figure or judgment | Caveat METR itself flags |
|---|---|---|
| Subjective pace (well-specified tasks, as of Oct 2026) | About 2× the March 2024–October 2025 rate | High uncertainty; 1.3× or 3× would not surprise them |
| AECI growth since October 2025 | 1.9× historical rate [95% CI 1.5–2.8×] | Could be benchmark overfitting or an AECI artifact |
| What “2× acceleration” means here | About two years of older progress in one year | Acceleration is a ratio of growth rates, not “twice as smart” |
| Robustness band across analysis choices | About 1.1×–2.9× | Sensitive to the statistical “link function” on benchmarks |
| Vs. the Haiku 5.5 system card | Card reported 1.7×; METR’s 1.9× is higher mainly because it includes newer unreleased models | Public cards and this review are not identical datasets |
| If the trend includes other labs’ frontier models | METR estimates a lower acceleration (one reading: ~1.4×) | Anthropic may have been sub-frontier in 2025, which would inflate a later catch-up |
| Training-compute explanation | Ruled out: acceleration was not from faster growth in training compute | Further compute details redacted |
| Share of current (Oct 2026) acceleration due to AI | METR’s guess: a majority, “perhaps ≈80%” | Substantial uncertainty; late-2025 speedup likely had more non-AI causes |
| Staff interviews on AI’s effect in their area | Centered around a 2× acceleration, with a wide range | Quotes partly redacted; people interpreted the question differently |
| Informal Slack-poll “uplift” | Reached about 12× for the most capable internal model in October 2026 | METR calls this very loose, low-effort, possibly unrepresentative evidence |
| Fully autonomous AI R&D | Not yet a large contributor to the most important training work | Researchers described “signs of life” that this could change soon |
| Near-term bottlenecks (next 12 months) | None clearly identified in algorithms, training compute, experiment compute, or data | METR’s view of bottlenecks changed several times as more context arrived |
AECI is Anthropic’s version of the Epoch Capabilities Index — a composite of benchmark scores, not a direct measure of real-world impact. METR dated models by when most Anthropic employees could use a snapshot, not by public release day, so it could compare unreleased systems on the same clock. The default trend break it estimates is October 7, 2025.
That last distinction matters. Public Claude releases can lag internal snapshots. Including three recent internal-only models is also why METR’s slope looks steeper than the number in the Haiku 5.5 system card.
Unreleased models in the dataset
The public report names the internal systems METR was allowed to discuss (these are the codenames shared with METR; the lab has different internal names):
- Unreleased 2022 — finished training in 2022.
- Model 2 — finished training in July 2026.
- Model 3 — finished training in August–September 2026.
- Model 4 — finished training in September–October 2026.
A forecast for a future model appeared in METR’s ~100-page full report and was redacted. Model 4 scores above 50% on every AECI benchmark it was tested on — a saturation warning METR takes seriously, because an index runs out of headroom when almost every task is already “easy.” Dropping three of 15 benchmarks whose score-vs-ability plots looked suspiciously flat or concave only raised Model 4’s AECI from 178.6 to 179.7, so METR does not think saturation alone explains the headline acceleration.
The Verge’s November 7 story highlights the same unreleased-model point: much of the productivity story METR describes is happening on systems the public cannot use yet. The newspaper separately notes staff self-reports of being around 12 times as productive with the most capable internal model, roughly double what staff had reported a few months earlier on Mythos 5. That 12× figure matches METR’s Slack-poll chart; both METR and The Verge treat self-reported uplift as suggestive, not as a clean productivity study.
How METR attributes the speedup to AI
METR separated three rates for any progress measure: the historical growth rate, the current growth rate, and a counterfactual rate if Anthropic could not use AI. Only the first two are directly observable. The third is a judgment.
After reviewing compute, hiring, data, algorithms, and other candidate causes, METR says it could confidently rule out faster training-compute scaling. Beyond that, even Anthropic staff did not have a single consensus story. METR’s synthesis is:
- The late-2025 acceleration likely had important non-AI contributors (details of many of those factors were redacted).
- AI’s contribution appears to have grown over time. METR’s October 2026 guess is that a majority — perhaps about 80% — of the current acceleration is due to AI.
- Most of the present speedup still looks like AI assisting human researchers, not AI independently delivering the most important training results.
Interview notes published with sensitive details removed are colorful and inconsistent in the way real labs are. One researcher said they could “definitely endorse a greater than 2X speedup,” described an “80% reduction in time between idea and experiment,” and no longer felt bottlenecked on code. Another compared Claude to having postdocs and guessed they were about 3× more productive. Others credited hiring as the bigger story, or guessed only 15–20% of the change was AI. METR asked Claude to classify the interview set against a “greater than 2×” bar and got a mixed tally (two yes, three leans-yes, four unclear, two no) — presented as mildly supportive, not decisive.
A useful caution from the report, easy to miss in headlines: a person reporting “I am 12× faster” is not the same thing as the organization achieving a 12× faster rate of capability progress. Sustaining output growth usually requires sustained growth in labor or labor productivity just to hold the old pace against diminishing returns. METR cares more about whether individual uplift itself is still accelerating than about any one absolute multiplier.
What “no bottlenecks” does and does not mean
The risk METR is trying to illuminate is compounding: more capable models speed up R&D, which produces more capable models. If inputs (compute, data, algorithms, experiment capacity) hit a wall, that loop slows. METR did not see convincing evidence of such a wall in the next 12 months — and immediately adds that its understanding of specific bottlenecks flipped more than once as Anthropic shared more context, so it is not confident in the current assessment. Analysis of several specific bottlenecks was redacted.
METR also lists ways it could be wrong in either direction. Acceleration could be higher than advertised if growth is becoming hyperbolic, if AECI lags real R&D (you only see a trained model after months of work), if full automation is closer than it looks, or if real-world impacts jump when models cross thresholds that benchmarks move through smoothly. Acceleration could be lower if non-AI factors explain most of the change, if recent models are improving faster on clean, auto-gradable tasks than on messy ones, or if Anthropic’s 2023–2025 baseline was unusually slow compared with the wider industry frontier (which would make a later catch-up look like a permanent doubling).
The redactions are part of the story
METR delivered a roughly 100-page full report to Anthropic on October 22 and iterated on a public version through early November (the final public report was submitted November 3, after a short extension to handle tightened redaction guidance). Anthropic had unilateral rights to redact non-public information. METR retained the right to say that redaction happened, whether key findings depended on redacted material, and what level of access it received.
While we were able to freely share our highest-level conclusions about risk, Anthropic redacted a majority of the evidence we found most informative for these conclusions.
METR, November 5, 2026 public report
Anthropic described the cuts as aimed at confidential information that could accelerate other AI developers. METR agrees the redacted material was confidential and potentially useful to rivals, and notes that most of it was “very high-level.” The group does not try to decide whether the business interest outweighs the public interest; it focuses on what readers can no longer check for themselves.
A table in the report is blunt about the gap. METR saw internal capability measurements, algorithmic-progress estimates, input metrics, bottleneck details, and much of the evidence on AI’s contribution. The public version keeps high-level benchmark results and only part of the AI-contribution and interview evidence. Intermediate measures, algorithmic-progress estimates, most input metrics, bottleneck details, detailed acceleration-by-pipeline-stage numbers, and forecasts of future models are out. METR’s complaint is not theatrical: when evidence has “multiple plausible interpretations,” hiding the evidence forces the public to take METR’s judgment on faith.
How outside researchers and The Verge read it
The Verge’s November 7 article frames the report as some of the clearest public evidence yet that AI is speeding up its own development — and as arriving while Anthropic prepares to go public and while CEO Dario Amodei has argued that labs should pump the brakes. Hart quotes Anthropic cofounder Sam McCandlish’s lengthy response on X:
We redacted a lot, and I accept that this makes it harder for others to check METR’s work. But we determined that sharing exactly what is driving the speedup isn’t the right thing to do, both because it’s commercially sensitive and because it risks handing out a roadmap for others to go faster.
Sam McCandlish, as quoted by The Verge, November 7, 2026
McCandlish, again via The Verge, said sharing that roadmap would make the risks METR worries about worse, and that Anthropic plans to publish its own report on the speed of AI progress “in the coming weeks,” plus new safety steps. Anthropic declined to answer The Verge’s questions and pointed the paper to that post.
Safety researchers quoted by The Verge treated the pace as at least as fast as many insiders expected, and possibly faster than the well-known “AI 2027” scenario. Apollo Research founder Marius Hobbhahn told the paper the report offers evidence on something that has been hard to pin down: how quickly AI is accelerating its own development. Ryan Greenblatt, a METR researcher, wrote that he “was seriously wrong about RSI speed up; I previously expected much less in 2026.” Daniel Kokotajlo of the AI Futures Project accused Anthropic of “recklessly gambling with our lives” and called it “pretty rich” to cite rival speed as a reason to redact while not appearing to slow down much itself. Brianna Rosen of the Institute for AI Policy and Strategy argued that independent evaluators need freedom to explain their conclusions if democratic oversight is going to work.
The Verge also notes political reaction: New York Assembly member Alex Bores called on Congress to “ban RSI, this year if possible.” That is an advocacy statement, not a law, and it is not part of METR’s own recommendations. METR’s document is an assessment, not a legislative draft.
Separately, The Verge records that OpenAI said in September it is “making strong progress” toward an AI researcher that can “further progress on deep learning and alignment” by March 2028 — context that the acceleration question is not unique to Anthropic, even though this particular dataset is.
What this is — and what it is not
| Claim | Status | Notes |
|---|---|---|
| METR ran an 11-week high-access review at Anthropic (Aug 12–Oct 29, 2026) | Confirmed by METR | Distinct from the Oct 15 “employee-like access” evaluator program |
| Public write-up dated November 5, 2026; full internal report ~100 pages | Confirmed by METR | Final public draft submitted November 3 after redaction iterations |
| About 2× faster progress on well-specified tasks vs. March 2024–October 2025 | METR’s subjective conclusion | Uncertainty band roughly 1.3×–3× |
| AECI trend ≈1.9× since October 2025 [1.5–2.8×] | METR statistical estimate | Robustness checks span about 1.1×–2.9× |
| Majority of current acceleration due to AI (guess ≈80%) | METR judgment | Late-2025 speedup likely mixed; details redacted |
| Not caused by faster training-compute scaling | METR says it is confident | Underlying compute evidence redacted |
| Staff Slack-poll uplift ≈12× in October 2026 | Reported by METR (and The Verge) | METR: loose, low-effort self-reports |
| Fully autonomous AI scientist already running the training stack | Not what METR found | “Signs of life” for more autonomy later |
| Clear 12-month bottlenecks identified | No | METR is explicitly uncertain |
| The public can inspect the most informative evidence | No | Anthropic redacted most of it |
FAQ
Did METR say Anthropic’s AI is twice as capable as last year’s?
No. It said the rate of improvement on well-specified tasks looks about twice what it was between March 2024 and October 2025. A 2× acceleration means roughly two years of the old pace packed into one year, not that today’s model is “twice as smart.”
Is the 12× productivity number a formal time-and-motion study?
No. It comes from informal Slack polls in which staff reacted with their guessed speedup from the latest internal model. METR repeatedly warns that the responses are low-effort and may not represent the whole company, and that survey uplift often overstates true productivity.
Why does this matter if most of the work is still humans using AI assistants?
Because assistant speedups can still change the slope of research, and because METR heard “signs of life” for more autonomous contributions. The worry in METR’s framing is compounding: faster models produce faster research, which produces faster models, especially if no input bottleneck appears.
What exactly did Anthropic hide?
METR cannot list every redaction without revealing the underlying secrets. In broad terms: detailed acceleration numbers for stages of the training pipeline, most intermediate metrics, algorithmic-progress estimates, most input measurements, bottleneck analyses, future-model forecasts, and a large share of the evidence that most informed METR’s risk conclusions.
Is this the same review as Anthropic’s October 15 independent-evaluator program?
No. METR says this project started in mid-August, was never intended to include full employee-like access, and is distinct from the October 15 announcement. Reviewers did not query unreleased models.
Where should I read the primary sources?
Start with METR’s November 5 public report and then The Verge’s November 7 analysis for how outside safety researchers and Anthropic’s Sam McCandlish responded. Treat social-media summaries that add extra benchmark-hour figures as unverified unless those figures appear in the report itself.
Sources: METR — Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 (November 5, 2026; AEF-1 checklist corrections logged November 6); The Verge — Why a new report on Anthropic automating AI development is alarming researchers (Robert Hart, November 7, 2026, 12:30 p.m. UTC). Figures and quotations above are taken from those two pages. We did not independently audit Anthropic’s internal data, and we do not treat redacted evidence as if it had been published.