METR Says Anthropic Is Improving Its Models About Twice as Fast, Mostly Because of AI. Anthropic Redacted Most of the Evidence the Public Would Need to Check That.
METR published its independent assessment on November 5 after about eleven weeks inside Anthropic. Its judgment: as of October 2026, Anthropic is likely improving capabilities on well-specified tasks at around twice its March 2024 to October 2025 rate, with a majority of the current acceleration, METR guesses about 80%, coming from AI use, and no clear bottleneck in the next twelve months. The measured figure is a 1.9x acceleration on Anthropic's capability index since October 2025, with a 95% confidence interval of 1.5x to 2.8x and robustness runs spanning 1.1x to 2.9x. Anthropic redacted a majority of the evidence METR found most informative. We read METR's post.
By T.W. Ghost
METR published today, November 5, 2026, the results of about eleven weeks embedded at Anthropic, from August 12 to October 29, with roughly two and a half full-time staff, unpaid, working on Anthropic devices. The question was whether AI is speeding up Anthropic's own AI research. METR's answer, stated as a judgment and not a measurement: as of October 2026, Anthropic is likely improving capabilities on well-specified tasks at around twice the pace it managed from March 2024 to October 2025. METR says it would not be surprised by 1.3x or by 3x. Its guess is that a majority of the current acceleration, around 80%, is due to AI use, and that it sees no clear bottleneck in algorithms, training compute, experiment compute, or data that would stop the compounding over the next twelve months, with the caveat that it is not confident about that either.
The measured number underneath is an acceleration of 1.9x since October 2025 on Anthropic's variant of the Epoch Capabilities Index, with a 95% confidence interval of 1.5x to 2.8x and a breakpoint around October 7, 2025. Robustness checks with different link functions range from 1.1x to 2.9x; most other specifications land between 1.9x and 2.2x. And then the sentence that frames the whole report: Anthropic redacted a majority of the evidence METR found most informative in the roughly 100-page full version, including intermediate capability metrics, algorithmic-progress estimates, most input metrics, bottleneck details, and forecasts. METR says so plainly and says the redaction fails its standard of transparency about evidence. We read METR's post. Here is what held up.
What METR Found, and What You Can See
| METR's assessment of Anthropic (November 5, 2026) | |
|---|---|
| Engagement | August 12 to October 29, 2026; about 2.5 FTE; unpaid; Anthropic devices |
| Headline judgment | About 2x the historical pace on well-specified tasks; 1.3x to 3x would not surprise METR |
| Measured acceleration | 1.9x on Anthropic's capability index since October 2025; 95% CI 1.5x to 2.8x; breakpoint about October 7, 2025 |
| Robustness | 1.1x to 2.9x across link functions; mostly 1.9x to 2.2x otherwise |
| Source of acceleration | Not faster training-compute scaling (METR confident; details redacted); majority, about 80%, from AI use (METR's uncertain guess); late-2025 mix included non-AI factors |
| Staff self-reports | Slack polls reached about 12x self-reported uplift in October 2026; METR calls this very loose evidence |
| Autonomy | AI not yet making highly significant fully autonomous contributions to the most important training stages; "signs of life" |
| Bottlenecks | None clear for twelve months; METR not confident |
| Models in the fit | Includes unreleased internal models finishing July to October 2026 |
| Prior figure | Haiku 5.5's system card cited 1.7x; METR's 1.9x mainly adds newer internal-only models |
| Redacted | A majority of the most informative evidence: intermediate capability metrics, algorithmic-progress estimates, most input metrics, bottleneck details, forecasts |
Twice as fast is a band, not a point. The interval runs from 1.5x to 2.8x, the robustness runs from 1.1x to 2.9x, and an industry-wide framing of the same data can read closer to 1.4x. Anyone quoting "2x" owes the reader the band.
Twelve times is not two times. The Slack polls where staff report their own uplift hit about 12x. METR is explicit that individual self-reported uplift and organizational capability growth are different quantities, and it weights the polls accordingly.
The redaction is the story. The public gets METR's conclusions and a confidence interval on one index. It does not get the measurements that would let an outsider reconcile Anthropic's curve with another lab's, or test METR's attribution guess. METR's own words are that this fails on transparency; we have nothing to add to them.
How It Fits With the S-1
Anthropic's amended S-1, filed November 2, contains a risk factor saying internal estimates about future models indicate the company "may imminently meet" its Responsible Scaling Policy's definition of dramatically accelerated progress attributable to AI. METR's report and that filing are related and not the same thing. METR measured the pace of capability improvement and attributed most of the current acceleration to AI; the RSP threshold is Anthropic's own definition, applied by Anthropic, and no current model has been designated as over it. Putting the two side by side is fair. Treating either as proof of runaway recursive self-improvement is not, and METR's "signs of life" language on autonomous contributions points the other way for now.
This engagement is also distinct from Anthropic's October 15 announcement about independent evaluators, and a compute-allocation verification METR planned was canceled partway through.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Engagement dates, staffing, unpaid, Anthropic devices | Verified (METR post) |
| About 2x judgment; 1.3x to 3x plausible | METR's judgment |
| 1.9x acceleration, CI 1.5x to 2.8x, breakpoint about Oct 7, 2025 | METR on Anthropic's data (underlying data redacted) |
| Robustness 1.1x to 2.9x | METR |
| About 80% of acceleration from AI | METR's uncertain guess |
| About 12x self-reported uplift in staff polls | METR, citing Anthropic poll data; very loose evidence by METR's account |
| No clear bottleneck for twelve months | METR, not confident |
| Majority of most-informative evidence redacted by Anthropic | Verified (METR's disclosure) |
| Relationship to the S-1/A RSP risk factor | Related, not identical; different documents and definitions |
What To Do
Read METR's redaction table before you quote the 2x, and quote the interval with it.
Keep the METR number and the RSP risk factor in the same paragraph and different sentences when you write about Anthropic's IPO; one is an outside measurement with a band, the other is a company's own threshold language.
The frontier is elsewhere: the report is about how fast the models improve, not where they stand today, and today's standings moved again this week. Our model comparison has the current board side by side.
Sources
- METR: assessment of AI R&D acceleration at Anthropic
- AP: Anthropic, METR, and AI acceleration
- The Verge: an Anthropic risk assessment from AI safety organization METR is out
- SEC EDGAR: Anthropic PBC Form S-1/A No. 1 (RSP risk factor context)
*Not sure what a lab's rate of improvement should change about which model you pick today? Take the free 2-minute quiz and get matched. Then read the filing METR's report now sits beside: Anthropic's S-1.*