METR says AI progress at Anthropic has roughly doubled in speed, mostly thanks to AI itself

An 11-week review with Anthropic’s internal data

On Nov. 5, the nonprofit evaluation group METR published a report on a high-access assessment it ran at Anthropic to measure how much AI is accelerating AI research and development (R&D). From Aug. 12 to Oct. 29, 2.5 full-time-equivalent METR staff worked on Anthropic devices. Anthropic shared pretraining and reinforcement learning metrics, benchmark scores for internal models and training compute data, and METR interviewed about 15 staff. METR says it took no payment from Anthropic. The project is separate from the “employee-like access” assessment Anthropic announced on Oct. 15.

METR@METR_Evals

Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development.

We are now publishing a report summarizing the results of this risk assessment.

Three subjective conclusions

First, METR believes that as of October 2026, Anthropic is likely improving its models on well-specified tasks at around twice the rate it did from March 2024 to October 2025. It would not be surprised if the true figure were closer to 1.3x or 3x. Second, a majority of today’s acceleration, perhaps about 80%, is likely due to Anthropic’s heavy use of AI in R&D, though METR is not certain and Anthropic staff disagree. Third, METR saw no clear bottlenecks, so it sees a significant risk of even faster acceleration over the next 12 months.

Figures the report could share

• Anthropic’s AECI scores (its version of the Epoch Capabilities Index) have grown at 1.9x the earlier rate since October 2025 (95% confidence interval: 1.5x to 2.8x).
• Other analysis choices give 1.1x to 2.9x. Across all developers’ frontier models, the figure is 1.4x.
• Anthropic’s Haiku 5.5 system card reported 1.7x. METR’s higher figure mainly reflects unreleased internal models.
• In informal Slack polls, staff reported an average speedup of about 12x from the best internal model in October 2026. METR calls this “very loose evidence.”

METR is confident the acceleration did not come from faster scaling of training compute.

Redacted evidence and the risk report

METR gave Anthropic a roughly 100-page full report on Oct. 22. In the public version, Anthropic redacted a majority of the evidence METR found most informative. Anthropic said the redactions removed confidential information that could accelerate other AI developers. They include algorithmic progress estimates, bottleneck analyses and forecasts for a future model. METR says the public therefore has to rely on its judgment much more than would be ideal. Anthropic also canceled a planned check of how its compute is split between capabilities and safety research.

In a separate review of Anthropic’s August 2026 Risk Report, METR agrees that catastrophic risk from models as of July 15 was low. But it finds it plausible that the early warning threshold in Anthropic’s Responsible Scaling Policy had already been crossed. According to METR, indicators the public report called “less than a factor of 2” imply about 1.97x in the private version.