METR Estimates Anthropic’s Model Capability Growth Has Nearly Doubled

METR published an independent assessment on November 5, 2026, judging that Anthropic is likely improving model capabilities on well-specified tasks at around twice its earlier rate and that AI use likely explains a majority of the current acceleration. The conclusions carry substantial uncertainty, and redactions conceal much of the supporting evidence.

Benchmark Evidence Shows a Steeper Trend

Using AECI, Anthropic’s benchmark capability index, METR estimated that benchmark scores have grown 1.9 times faster since October 2025, with a 95% confidence interval of 1.5–2.8 times. That rate implies roughly two years of progress at the previous pace in a single year. The broader judgment compares October 2026 with the rate observed from March 2024 through October 2025.

The calculation includes internal-only models and dates each snapshot by when most Anthropic employees could use it, creating a common basis for released and unreleased systems. METR’s robustness checks produced acceleration estimates ranging from 1.1 to 2.9 times under alternative modeling choices, demonstrating sensitivity to how benchmark difficulty and other analytical assumptions are represented.

Evidence for AI’s Role in the Speedup

The evidence convinced METR that faster growth in training compute did not cause the acceleration, although the supporting details were redacted. Non-AI factors likely played a large role in late 2025, but the current best guess is that they do not explain the sustained speedup.

Staff evidence pointed in the same direction but was uneven. Researcher interviews varied widely around an approximately twofold acceleration. Informal polls put average self-reported uplift from the most capable internal model at about 12 times in October 2026; METR cautioned that the responses required little effort, might not represent staff broadly, and that individual productivity gains do not directly equal organization-wide capability growth.

The current effect comes mainly from AI assisting researchers. AI cannot yet make highly significant, fully autonomous contributions to the most important training work, although researchers described indications that such contributions could emerge soon, with maturity varying across areas.

Bottlenecks and Measurement Remain Uncertain

No convincing evidence emerged that algorithmic progress, training compute, experiment compute or data would impose significant bottlenecks over a 12-month horizon. The reviewers’ interpretation changed several times as new information arrived, leaving them unconfident in that assessment. METR nevertheless judged that the absence of clear constraints or steep diminishing returns creates a significant risk that further capability gains could compound into even faster acceleration; that is a forward-looking risk judgment, not an observed result.

AECI emphasizes well-specified, readily graded tasks, so faster benchmark gains may not translate into equivalent real-world effects. METR said overfitting, non-AI contributors or an unusually weak historical baseline could make its estimate too high, while missed recent acceleration, faster-than-expected removal of human bottlenecks or benchmark scores that understate real-world impact could make it too low. Model 4’s scores above 50% on all 15 benchmarks used for it added a separate concern that saturation could impede future measurement.

Extensive Access but Limited Public Evidence

METR staffed the eleven-week assessment at 2.5 full-time equivalents, running from August 12, 2026, through October 29, 2026. It reviewed internal benchmark results, pretraining and reinforcement-learning metrics, training-compute measurements, AI-speedup analyses and internal documents, and drew on roughly 15 interviews with key staff. The team did not receive unmediated, employee-like access or sample non-public models.

Anthropic could approve or reject information requests and unilaterally redact non-public material. Its redactions removed from the public report a majority of the evidence METR considered most informative, including detailed AI attribution, algorithmic-progress estimates, specific bottleneck analysis and forecasts for future models. Anthropic described the redactions as protection for confidential information that could accelerate other AI developers. METR retained discretion over methodology and conclusions and accepted no payment, though it worked on Anthropic devices and used free tokens.

The assessment provides an extensive outside examination of AI-driven research acceleration while stopping short of transparency about the underlying evidence. Public benchmark results and staff accounts support faster progress, but redactions prevent readers from independently assessing much of the basis for METR’s attribution and bottleneck conclusions. Those central findings therefore rest heavily on the evaluator’s synthesis of evidence that readers cannot fully inspect.