Metr assessment finds anthropic is building ai models twice as fast as last year, mostly thanks to ai

What did METR find at Anthropic?

From August 12 to October 29, 2026, the Model Evaluation and Threat Research group conducted a high-access evaluation at Anthropic, reviewing internal benchmark data, training-compute metrics, staff surveys, and AI speedup analyses. The goal was to answer a high-stakes question: is AI itself accelerating AI research and development?

METR's headline finding: 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. The estimate comes with significant uncertainty — the true acceleration could be closer to 1.3 times or as high as 3 times the historical rate, the evaluators wrote in their report published November 5.

Since October 2025, benchmark scores measured via Anthropic's version of the Epoch Capabilities Index (AECI) have grown at an estimated 1.9 times the pre-October 2025 rate, with a 95% confidence interval of 1.5 to 2.8 times. METR is confident this acceleration was not caused by an acceleration in training-compute scaling, ruling out the simplest explanation for why models are getting better faster.

How much of the speedup is from AI itself?

METR attributes a majority of the current acceleration — around 80% — to Anthropic's extensive use of AI in its own R&D process. Staff surveys and interviews are consistent with AI usage causing a doubling in the pace of capabilities progress, and Anthropic employees self-report dramatic productivity uplift from AI assistance.

However, METR notes that the acceleration in late 2025 was likely influenced by non-AI factors. AI's contribution has grown relatively more important over time. The acceleration is primarily from AI assistance rather than fully autonomous AI contributions, though METR says there are signs fully autonomous work will play a larger role soon.

METR found no evidence of significant bottlenecks that would prevent further acceleration from compounding — meaning as AI builds better AI, the cycle could continue to speed up. However, the team's understanding of bottlenecks changed several times as it received more detailed information, so METR says it is not confident in its current assessment.

What did Anthropic redact from the report?

Anthropic had the right to redact any non-public information in the report. It exercised that right extensively.METR says Anthropic redacted a majority of the evidence the evaluators found most informative for their conclusions, describing the redactions as aimed at removing confidential information that could accelerate other AI developers. METR shared a table of what it was and was not able to discuss, but much of the redacted evidence had multiple plausible interpretations, with both METR and Anthropic researchers expressing significant uncertainty.

The lack of public access to key evidence means readers must rely on METR's judgment far more than would be ideal for questions of this weight.As noted in METR's earlier assessments of OpenAI's agent monitoring, the pattern has been consistent: independent evaluators consistently see more than they can publish, leaving the public to trust their conclusions without access to the underlying evidence that shaped them.

What does this mean for AI safety and governance?

The finding that AI is measurably accelerating AI R&D has direct implications for how fast frontier capabilities could advance. METR previously published a framework asking whether the public can judge loss-of-control risks from frontier AI development without transparency into internal development metrics. This assessment is the first attempt to answer that question with direct access.

Anthropic offered unprecedented access for this evaluation — far more than any other frontier lab has provided to independent evaluators. Yet the scope of redactions means the transparency is incomplete. For the public, the bottom line is that METR's team — having seen the data — believes the acceleration is real and significant, but cannot share the evidence that convinced them.

Previous assessments have drawn similar conclusions about the limits of independent evaluation.METR's agreement with Anthropic on AI risk monitoring also found key evidence was withheld, raising the same question: how much transparency is enough when the stakes involve potential loss of control over frontier AI systems?

How reliable is the acceleration estimate?

METR tested the robustness of its 1.9 times acceleration estimate across different choices of benchmark inclusion, model dating, model inclusion, and benchmark saturation. Robustness analyses produced acceleration rates between 1.1 and 2.9 times. METR says it would be surprised if the true evolution showed no acceleration or acceleration above 3 times.

Potential overestimates could come from non-AI factors being more significant than assumed, AECI's focus on verifiable tasks missing harder-to-measure real-world capability gaps, or Anthropic's earlier growth rate (2023-2025) being unusually low, making the current pace look faster by comparison.

Potential underestimates could come from recent growth rates that may have further increased, increasing automation of AI R&D tasks, or AECI growth historically understating real-world impact growth. METR staff remain internally divided on some of these questions, the report notes.

What comes next?

The METR assessment arrives as Anthropic is navigating the most consequential period in its history.The company filed for an IPO in October and is reportedly targeting a valuation of $1.2 to $1.3 trillion, down from earlier estimates of $2 trillion. Investors are weighing whether AI companies can sustain the growth rates the market has priced in.

If AI is genuinely accelerating AI R&D at the rate METR estimates, the implications extend beyond Anthropic. The question applies to every frontier lab: when AI makes AI development faster, what limits the feedback loop? METR's inability to identify clear bottlenecks means the risk of compounding acceleration is real, even if the team's confidence in that conclusion is limited by the data it could not share.

The full 100-page METR report is available on the organization's website. A shorter announcement thread is on X from the @METR_Evals account.