METR published an independent assessment on November 5 concluding that Anthropic is likely improving the capabilities of its models on well-specified tasks at around twice the rate it did between March 2024 and October 2025. METR guesses that a majority of that speedup, perhaps about 80%, comes from Anthropic's extensive use of AI to do its own R&D.
The findings come from a high-access evaluation METR ran inside Anthropic from August 12 to October 29, 2026, with 2.5 full-time staff working on Anthropic devices. Anthropic supplied free tokens, and METR says it took no payment. The work set out to answer three questions: whether Anthropic's internal capability growth has accelerated, how much of that is attributable to AI, and whether any bottlenecks could slow growth in the near term.
The core metric is AECI, Anthropic's version of the Epoch Capabilities Index, which tracks benchmark scores across Anthropic models. METR measured a 1.9x acceleration since October 2025, with a 95% confidence interval of 1.5x to 2.8x, while alternative fits produced estimates between 1.1x and 2.9x. Put another way, Anthropic made as much benchmark progress in the past year as it did between October 2023 and October 2025, roughly the span from Claude 2.1 to Sonnet 4.5.
Key facts
- 01METR measured a 1.9x acceleration in Anthropic's capability growth since October 2025 (95% CI 1.5x to 2.8x).
- 02METR guesses a majority, perhaps about 80%, of the current acceleration is due to AI.
- 03The evaluation ran from August 12 to October 29, 2026, with 2.5 full-time METR staff.
- 04Anthropic's own Haiku 5.5 system card estimated a 1.7x acceleration.
- 05Anthropic redacted a majority of the evidence METR found most informative.
That figure is higher than the 1.7x estimate in Anthropic's own Haiku 5.5 system card. METR attributes the gap primarily to data on newer unreleased models, which it labels Model 2 (July 2026), Model 3 (August to September 2026) and Model 4 (September to October 2026). METR says it did not request or receive access to sample from non-public models.
Anthropic shared internal benchmark scores, pretraining and reinforcement learning efficiency history, headcount, training compute, internal surveys of self-reported AI uplift, lines of code and pull requests, AI agent inference spend, internal documents on autonomous research, and about 15 interviews with key staff.
METR says it is confident the acceleration was not driven by faster growth in training compute. It describes Anthropic's R&D as accelerated primarily through AI assistance rather than fully autonomous contributions, and says AI is not yet able to make highly significant, fully autonomous contributions, though it cites signs of life on that front. In internal Slack polls, which METR calls very loose evidence, self-reported uplift rose from about 1.2x for Claude 3.7 Sonnet to about 12x for Model 4.
On bottlenecks, METR reports no evidence of significant constraints in algorithmic progress, training compute, experiment compute or data that would take effect in the next 12 months.
METR flags high uncertainty and says it would not be surprised if the true acceleration were closer to 1.3x or 3x. Factors that could mean an overestimate include non-AI causes explaining most of the gain and the acceleration being limited to well-specified tasks. Factors pointing the other way include AECI understating real-world capability impact.
Anthropic held unilateral redaction rights over non-public information and, per METR, redacted a majority of the evidence METR found most informative. Anthropic said the redactions were aimed at removing confidential information that could accelerate other AI developers. Anthropic also narrowed the project midway to exclude compute verification, saying it wanted to keep the engagement focused on risk assessment and would revisit compute allocation verification later.
METR says the engagement is separate from the assessment Anthropic announced on October 15, and that it submitted its final public report on November 3.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
