Key takeaways
- METR published an independent assessment on November 5 saying Anthropic improves capabilities on well-specified tasks at about 2x its March 2024 to October 2025 pace.
- METR estimates about 80% of the acceleration is due to AI, a subjective judgment with substantial uncertainty.
- Anthropic redacted a majority of the evidence METR found most informative, so readers must rely on METR's judgment.
METR, the nonprofit that tests AI systems, says Anthropic is improving its AI models about twice as fast as it did before October 2025, and that AI itself drives roughly 80% of the speedup. The finding comes from METR's independent assessment of AI R&D acceleration at Anthropic, published on November 5.
The report matters because Anthropic's own safety policy has a threshold for exactly this: AI speeding up AI research by 2x. METR also published a separate review of the automated R&D section of Anthropic's August 2026 Risk Report. It says that report understates the acceleration.
What METR found
METR staff worked with Anthropic's internal data from August 12 to October 29. They conclude that, as of October 2026, Anthropic is likely improving capabilities on well-specified tasks at about 2x the rate of March 2024 to October 2025. They say a figure between 1.3x and 3x would not surprise them, and that anything above 3x would.
The main benchmark evidence is AECI, Anthropic's version of Epoch's Capabilities Index. Its trend rose from about 15 points a year before October 2025 to about 29 points a year after, an acceleration METR puts at 1.9x (95% confidence interval 1.5x to 2.8x). Frontier models from all developers show only about 1.4x over the same period. METR's robustness checks range from 1.1x to 2.9x.
On the cause, METR estimates about 80% of the October 2026 acceleration is due to AI. It calls that a subjective judgment with substantial uncertainty. It says it is confident the speedup did not come from faster growth in training compute.
What the review says about Anthropic's Risk Report
Anthropic's public August report says acceleration is "less than a factor of 2." METR says the private version's figures imply about 1.97x, and that its own analysis gives 1.3x to 2.8x depending on definition and period. METR thinks it is plausible that models as of July 15 had crossed the 2x early warning threshold in Anthropic's Responsible Scaling Policy, though most researchers on the project and at Anthropic lean toward not.
It agrees that catastrophic risk from those models was low, and finds the report's argument that the Full Substitution Threshold was not crossed broadly compelling, with lower confidence than in past reports. METR cannot rule out substantial further acceleration in the next six months, which could lead to a "high" risk rating.
What METR could not show
Anthropic had unilateral rights to redact the report. METR says it removed "a majority of the evidence" METR found most informative, including intermediate capability measures, algorithmic progress estimates and forecasts for future models. Anthropic's stated reason was removing confidential information that could accelerate other AI developers. METR says the redactions did not change its highest-level conclusions, but readers must rely on its judgment more than it considers ideal.
Anthropic also dropped a contracted check of how it splits compute between capabilities and safety, saying it wanted to keep the engagement focused on risk assessment for now.
Other limits METR lists:
- It had no staff-level access and no query access to unreleased models.
- Anthropic's most recent model scores above 50% on all 15 AECI benchmarks, so the benchmark set is nearing saturation.
- Staff self-reports of productivity uplift rose from about 1.2x to about 12x across model generations, which METR calls "very loose evidence."
- METR found AI is not yet making highly significant, fully autonomous contributions to the most important training work, though researchers reported "signs of life."
- One METR team member had discussed a possible Anthropic role and held equity in a competitor as a former employee. A colleague without a conflict checked that person's work.
What to watch
Anthropic's own account of the findings was not available when this article was published. Watch for it, and for whether the company declassifies any of the redacted evidence. Our earlier coverage of the company's IPO filing covers its own warning about self-building AI.
