METR assesses that Anthropic speeds up its AI R&D by ~2x, but with most of the evidence redacted

Editor's note: the content received is a partial version of the METR report (it cuts off in the process-transparency appendices). This summary is based solely on what appears in that text and does not cover the full report or its appendices.

METR, an independent evaluation organization, published on 5 November 2026 an elevated-access evaluation of the extent to which AI is accelerating Anthropic's own AI R&D. The work was carried out between 12 August and 29 October 2026, with METR staff equivalent to 2.5 full-time people over eleven weeks. Anthropic provided access to internal data on its R&D process: pretraining and reinforcement learning (RL) metrics, training compute by domain, internal model benchmark results, speedup analyses, metric forecasts for a future model, and staff interviews.

A central caveat in the report itself concerns redaction. Anthropic had the right to redact any non-public information, and according to METR it removed most of the evidence METR considered most informative. Anthropic justified the cuts as aimed at removing confidential information that could accelerate other AI developers. METR acknowledges that some of the redacted evidence allowed several plausible interpretations and that both its own researchers and Anthropic's have considerable uncertainty about how to synthesize it. It therefore concedes that the report does not achieve "transparency about the evidence" and that the public must rely on METR's judgment more than would be desirable. METR retained the right to disclose that redactions occurred and whether its conclusions depended on redacted material.

Main conclusion: as of October 2026, Anthropic is probably improving its models' capabilities on well-specified tasks at roughly twice the rate it achieved between March 2024 and October 2025. Uncertainty is high: METR would not be surprised if the real speedup were closer to 1.3x or 3x. There is further uncertainty about how that speedup carries over to broader, less well-defined tasks.

The key quantitative measure is AECI, Anthropic's version of the Epoch Capabilities Index. Since October 2025, Anthropic's model scores have been growing at 1.9 times the previous rate (95% confidence interval: 1.5–2.8x). Put differently, in the past year the company would have made as much benchmark progress as between October 2023 and October 2025 (roughly, from Claude 2.1 to Sonnet 4.5). This figure exceeds the 1.7x cited in the Haiku 5.5 system card, mainly because METR included data from newer, unreleased models. Robustness analyses yield speedups between 1.1x and 2.9x, and METR would be surprised if the true trajectory showed no acceleration or more than 3x. The team itself cautions that this figure could reflect overfitting to the benchmark distribution or be an artifact of the AECI methodology, and that the data are sparse and lag behind the actual acceleration of research.

On attribution to AI: METR determined with confidence that the acceleration is not due to faster growth in training compute, although the details of that evidence were redacted. Beyond that, there was no consensus among Anthropic staff and no decisive evidence. METR's subjective estimate is that the late-2025 acceleration was probably driven by non-AI factors, while a majority (perhaps around 80%) of the current acceleration is due to AI, with substantial uncertainty. They believe AI's contribution has grown over time. Detailed estimates of cumulative and instantaneous effects across the different stages of training were redacted.

Interviews and surveys are consistent with AI use doubling the pace of progress. Researchers gave a wide range of estimates, centered on a 2x speedup. In addition, internal Slack surveys showed rising self-reported productivity gains, at very high levels (up to 12x in October 2026). METR considers this evidence very loose, since the responses were not carefully considered, and stresses that individual speedup does not equal acceleration of overall progress: sustaining a growth rate in the face of diminishing returns requires similar growth in inputs, and what matters is the change in the growth rate, not the absolute level.

On autonomy, METR was told that R&D is being accelerated mainly through AI assistance rather than fully autonomous contributions, and that AI does not yet make very significant, autonomous contributions to the most important parts of training. However, researchers described "signs of life" that such contributions could arrive soon, with maturity varying widely by area.

On bottlenecks, METR saw no evidence of significant constraints in algorithmic progress, training compute, experiment compute or data that would bite within the next 12 months. But its understanding changed several times as it received more detailed information, so it has little confidence in that assessment, and Anthropic redacted the analysis of several specific bottlenecks. From this, METR concludes there is a significant risk that further capability gains will lead to even greater acceleration over the coming year. It also reviewed internal metrics on pretraining, RL, compute utilization, and inference and data efficiency; that evidence, hard to interpret and partly indicative of a speedup above 2x and rising, was its largest source of uncertainty.

Acknowledged limitations. METR could be underestimating the speedup if it gave too little weight to some growth rates having increased even more recently, to greater automation of R&D tasks accelerating progress further, or to AECI having at times undervalued real-world impact. It could be overestimating it if it misweighted non-AI factors, AECI's focus on verifiable tasks, or the reasons why Anthropic's early growth (2023–2025) was unusually low.

Process details. Anthropic includes AI R&D acceleration as a key risk indicator in its Responsible Scaling Policy (RSP), but METR says it did not focus on those thresholds, taking a holistic view instead. The engagement is separate from the evaluation Anthropic announced on 15 October. METR also reviewed the AI R&D section of Anthropic's August 2026 Risk Report, which it is publishing separately. Anthropic could unilaterally approve or reject METR's data requests; in practice, these were negotiated toward data that were informative, not extremely sensitive and easy to share, though METR says access exceeded its expectations. Access to sample non-public models was neither requested nor granted, nor was it full "employee-level" access. METR worked on Anthropic devices for security reasons, used free tokens provided by Anthropic mainly for AI-assisted data analysis and visualization, and states that, as usual, it accepted no payment. The text received breaks off before the conflict-of-interest disclosure and the details of the redacted-evidence table, so we cannot describe them.