METR's Anthropic acceleration assessment, published November 5, 2026, concludes that as of October 2026 Anthropic is likely improving model capabilities on well specified tasks at about twice the rate it did between March 2024 and October 2025. Most of that current speedup, METR guesses, comes from Anthropic's heavy use of AI in its own R&D, not from faster growth in training compute. The nonprofit spent eleven weeks inside Anthropic's data from August 12 to October 29, then published a public summary after Anthropic redacted much of the evidence METR found most informative.
This guide walks through what METR actually measured, how the 1.9x AECI figure works, how the finding sits next to Dario Amodei's call to pace the frontier, and what the redactions leave open. Facts are checked against METR's report, Reuters coverage of the same day, Anthropic's October 15 update on independent evaluators, and Anthropic's own measurements of AI paced development, as of November 6, 2026.
METR Anthropic acceleration findings at a glance
| Claim | What METR says | Confidence / caveat |
|---|---|---|
| Pace of progress | About 2x the March 2024 to October 2025 rate on well specified tasks | High uncertainty; 1.3x to 3x would not surprise them |
| AECI trend | 1.9x faster since October 2025 (95% CI 1.5x to 2.8x) | Robustness checks span about 1.1x to 2.9x |
| Cause | Not faster training compute growth; majority of current acceleration likely from AI use in R&D (rough guess ~80%) | Staff disagree; late 2025 had non AI factors |
| Autonomy today | Mostly AI assistance, not fully autonomous R&D contributions | "Signs of life" for more autonomy soon |
| Near term bottlenecks | No clear blockers in algorithms, training compute, experiment compute or data over the next 12 months | Understanding of bottlenecks shifted often; low confidence |
| Redactions | Anthropic redacted a majority of the evidence METR found most informative | METR could still publish its top level risk conclusions |
The short version: Anthropic appears to be making roughly two years of historical benchmark style progress in about one year, AI tools inside the lab are the leading explanation METR gives for the current rate, and the public cannot see most of the internal charts that shaped that judgment.
What AECI 1.9x actually means
AECI is Anthropic's version of the Epoch Capabilities Index, a score that combines many internal and public style benchmarks into one capability index. Epoch AI publishes the public ECI methodology that AECI builds on. METR used Epoch's public code, Anthropic's internal scores, and ordinary least squares with a single estimated breakpoint around October 7, 2025.
"1.9x acceleration" means the slope of AECI after that breakpoint is about 1.9 times the slope before it. In METR's wording, Anthropic has made about as much AECI progress in the past year as it made from roughly Claude 2.1 era models to Sonnet 4.5 over the prior two years. The confidence interval is 1.5x to 2.8x. Changing the statistical link function moves the estimate from about 1.1x to 2.9x, so the exact number is fragile even when the direction is not.
METR also included unreleased internal models (codenamed Model 2, Model 3 and Model 4 in the public writeup). Model 4 scored above 50% on every AECI benchmark it was tested on. That saturation is one reason METR worries about measuring future acceleration with the same basket of tasks.
Importantly, METR says it is confident the speedup is not from training compute growing faster. Further compute details were redacted.
Is AI causing Anthropic's faster progress?
METR separates three ideas that headlines often blur:
- Historical growth rate before the late 2025 breakpoint.
- Current growth rate as of October 2026.
- Counterfactual growth rate if Anthropic could not use AI at all (guessed, not observed).
Staff interviews and surveys center around a roughly 2x acceleration from AI use in many research areas, with a wide spread. Slack style self reports of personal uplift reached about 12x for the most capable internal model in October 2026. METR treats those polls as weak evidence, because people overstate productivity gains and the sample is not careful.
METR's own subjective guess is that a majority of the current (October 2026) acceleration, perhaps about 80%, comes from AI. It also says non AI factors likely mattered a lot in late 2025, and Anthropic staff do not all agree on the split. The detailed attribution tables were redacted.
On autonomy, METR was told that R&D is sped up mainly by AI assistance rather than fully autonomous contributions to the most important training stages. Researchers described "signs of life" that more autonomous work could matter soon. That lines up with Anthropic's public August 2026 snapshot, which said Claude "leads" 26% of AI R&D work on an Epoch style automation scale, with more than 90% of work at "AI collaborates" or higher, and none fully autonomous.
How this relates to Pace the Frontier
In September 2026, Anthropic CEO Dario Amodei argued that AI driven recursive self improvement was already speeding capability gains and that labs should deliberately pace the frontier. He named METR as the kind of outside group that should get employee like access.
Two separate METR tracks then followed:
- October 15, 2026: Anthropic said it had started giving six METR staff employee like access for ongoing risk assessment, with desks, hardware and broad internal tools, under the White House Accord on Super Intelligence. That program is what Amodei's essay described.
- August 12 to October 29, 2026: The acceleration assessment covered here. METR states this engagement is distinct from the October 15 embedded evaluator program. Access was high and surpassed METR's expectations, but it was not full unmediated employee like access, and Anthropic could approve or reject information requests.
So the November 5 report is an early public stress test of independent review, not the final form of embedded evaluation. It also lands awkwardly next to the pacing message: Amodei called for slowing capability gains, while METR's public numbers say Anthropic's measured progress rate roughly doubled.
Anthropic cofounder Sam McCandlish told reporters the findings broadly match his own view that AI is speeding Anthropic's research, that exact measurement is hard, and that Anthropic plans to share its own analysis of progress speed. Reuters and AP both carried that framing on November 5.
How this differs from Mythos autonomy claims
Anthropic's Mythos 5.1 system card said autonomy threat model 2 (AI that could fully automate or dramatically accelerate large top tier research teams) was not crossed, partly because there was no sustained AI driven doubling in the pace of progress and the model was not close to replacing senior staff. Outside testers at METR were described as consistent with that at the time. Our guide on how autonomous frontier AI models are covers those system card lines.
The November METR report is measuring something related but not identical: the organization's overall rate of capability growth on well specified tasks, and how much of the recent speedup looks AI driven. METR still says fully autonomous contributions to the most important training stages are not yet a large share. Readers should not treat "2x AECI slope" as proof that Mythos already replaced senior researchers.
OpenAI tracks a related idea as "AI self improvement" in its Preparedness Framework. For how the lab rulebooks compare, see Preparedness Framework vs Responsible Scaling Policy. Anthropic lists AI R&D acceleration as a key risk indicator in its RSP; METR says it did not grade Anthropic against those exact thresholds in this assessment.
What was redacted, and why it matters
Anthropic had unilateral rights to redact non public information. METR could disclose that redactions happened, whether findings depended on redacted evidence, and the level of access. Anthropic described the cuts as removing confidential information that could accelerate other AI developers.
METR's public table says it could describe internal benchmark capabilities, but not key intermediate measures, algorithmic progress estimates, most bottleneck details, future model forecasts, or its fuller AI attribution breakdowns. METR says some of the redacted evidence made the ~2x story look more serious, and some suggested acceleration could be above 2x and still rising.
The practical effect is that outsiders must lean on METR's judgment more than on inspectable evidence. METR is blunt about that limit. It also notes Anthropic canceled a planned compute allocation verification mid project, so this report does not audit how much compute goes to safety versus capabilities. Anthropic has published its own compute and automation snapshots separately.
Pros and cons of reading the METR report as decisive
Strengths
- Eleven weeks of high access to internal metrics, interviews and unreleased model scores.
- Clear top line: not compute scaling; likely AI assisted R&D; roughly 2x on well specified tasks.
- Explicit uncertainty ranges and robustness checks.
- Independent nonprofit; METR took no payment (tokens and devices only).
Limits
- Majority of the most informative evidence is redacted.
- AECI may track clean, graded tasks better than messy real world research impact.
- Historical Anthropic AECI growth may have been below the wider industry frontier, which could inflate the acceleration ratio.
- Access was mediated; METR's views of bottlenecks flipped several times as more context arrived.
- Self reported uplift polls are soft evidence.
What it means if you use Claude or follow AI risk
For everyday Claude users, nothing in the report changes chat plans or prices. The report is about Anthropic's internal research pace, not a new consumer model launch.
For buyers, policymakers and safety readers, three takeaways matter:
- Independent review can publish hard numbers, even when the lab redacts the charts behind them.
- AI assisted R&D is no longer a future scenario in METR's account of one leading lab; the open question is how fast assistance turns into autonomy.
- Pacing proposals and acceleration measurements now sit side by side. Anyone arguing for slower releases needs a story that fits a measured ~2x capability slope inside a lab that also argues for pacing.
Groups such as Redwood Research have worked with METR on other 2026 incident reviews. ChatGPT and other labs face the same recursive progress question under different frameworks; OpenAI's category is AI self improvement, as our ChatGPT safety context and the preparedness comparison explain.
FAQ
What did METR find about Anthropic AI acceleration?
As of October 2026, METR believes Anthropic is likely improving capabilities on well specified tasks at about twice the rate seen from March 2024 to October 2025. AECI scores have grown about 1.9x faster since October 2025. METR guesses most of the current speedup comes from AI use in R&D, not from faster training compute growth.
Is Anthropic's AI already recursively self improving?
Not in the strong sense of fully autonomous successor building. METR says acceleration today is mainly AI assistance to human researchers, with early signs that more autonomous contributions could grow. Anthropic's own August 2026 automation index likewise reported no fully autonomous AI R&D subset.
Why did Anthropic redact so much of the METR report?
Anthropic said the redactions remove confidential information that could help other AI developers go faster. METR agrees the cuts were limited to that kind of material, but notes that most of the evidence it found most informative was removed, so the public must trust METR's synthesis more than raw exhibits.
Does this contradict Anthropic's call to pace the frontier?
It creates tension rather than a formal contradiction. Amodei argued that AI driven speedup is real and that labs should slow capability gains so safety can catch up. METR's numbers support the claim that speedup is real at Anthropic. They do not, by themselves, say whether Anthropic is actually slowing releases.
Is the METR assessment the same as the October 15 embedded evaluator program?
No. METR states the August to October acceleration study is distinct from the employee like access program Anthropic announced on October 15, 2026. The study had unusually deep data access, but not the full embedded setup described in that announcement.
Should I stop using Claude because of this report?
The report does not describe a new product defect or a consumer safety recall. It is evidence about how fast Anthropic's internal models are improving and how much AI helps that work. Use the same caution you would with any frontier assistant: check sources, limit sensitive data, and follow lab usage policies.


