At Anthropic, AI is already accelerating model research

METR estimates that Anthropic is now progressing twice as fast on well-specified tasks since October 2025. The study mainly highlights AI’s effect on research, but redactions hide much of the evidence.

METR published on November 5, 2026, an independent evaluation conducted at Anthropic: according to this report, the company is now improving the capabilities of its models on well-specified tasks at a rate close to twice that observed between March 2024 and October 2025. The organization attributes most of this acceleration to the use of AI in research and development, while emphasizing significant uncertainty. The issue goes far beyond Anthropic's case, as it concerns the speed at which leading labs can advance their systems. And here, the dials are moving fast.

A measured acceleration, but far from certain

The core of the analysis relies on AECI, Anthropic's internal index derived from the Epoch Capabilities Index. METR estimates that benchmark scores have progressed 1.9 times faster since October 2025, with a 95% confidence interval ranging from 1.5 to 2.8 times. In other words, the team sees the equivalent of about two years of historical progress in one year. This reading compares the recent period to the trajectory observed between March 2024 and October 2025.

The sensitive point, however, is the robustness of the measurement. According to METR, results still vary between 1.1x and 2.9x depending on assumptions made about benchmarks, the statistical function used, or model selection. Some data also concerns unpublished internal models, dated at the time when most Anthropic employees could use them. This helps compare systems but also complicates external interpretation.

Why METR believes AI does most of the work

The most striking conclusion is this: METR judges that the increase in training compute speed does not explain the observed acceleration. The report therefore excludes a simple "more compute, therefore faster" effect. Instead, investigators believe that most of the current acceleration comes from the use of AI itself, mainly as an assistant to researchers.

Interviews conducted with staff go in the same direction, even if responses vary. Some researchers describe a productivity gain greater than 2x, others a more modest or less clear effect. METR also notes internal surveys where the perceived magnitude of the gain rises up to about 12x in October 2026 for the most capable internal model. But these self-assessments are fragile: they do not directly measure collective performance and often respond in the moment. In short, it is not a perfect thermometer.

Another closely monitored signal: partial automation of research tasks. Interviewed researchers say AI already helps with coding, running experiments, or reducing friction, without fully replacing human contributions in the most important pipeline stages. METR nevertheless notes “signs of life” that could herald fully autonomous contributions earlier than expected.

What redactions still prevent from verifying

The report is not read as a fully open file. Anthropic exercised its redaction rights on most of the elements METR considered most informative. Removed passages notably concern detailed estimates of AI attribution, measures of algorithmic progress, analysis of certain bottlenecks, and forecasts on future models. Anthropic says it wanted to remove information “aimed at removing confidential information that could accelerate other AI developers.”

This situation directly affects public interpretation. METR says it can share its high-level conclusions but acknowledges that transparency on evidence is incomplete. In practice, the public must therefore give more weight to the evaluator's judgment than it would like. This is a classic blind spot of audits of closed models: one sees the silhouette, not always the detail of the seams.

Measurement limits that matter for the future

METR also insists on the limits of its own method. AECI mainly measures well-specified tasks, easy to score automatically or with a language model, which may underestimate or overestimate the real effect on the concrete world. If models cross useful thresholds faster than their scores show, the real impact may grow more sharply than expected. Conversely, if benchmarks mostly capture formal skills, they may exaggerate what happens in more “messy,” more human tasks.

The report also signals an important tension: Anthropic's models now score over 50% across all 15 benchmarks used for Model 4. This reduces the margin for interpretation of some tests and may complicate future measurements. Fundamentally, METR does not just say Anthropic is accelerating. It mainly says this acceleration could still increase if AI tools continue to lighten research itself.

What the investigation covered at Anthropic

The evaluation took place from August 12 to October 29, 2026, over eleven weeks, with the equivalent of 2.5 full-time positions on the METR side. The team says it had access to internal metrics, benchmark results, training compute data, speedup analyses, internal documents, and about fifteen interviews. Anthropic did not provide employee-level access, but the level of openness exceeded what the team initially expected.

The working environment was not trivial either. Researchers worked on devices provided by Anthropic, with free token credits, without direct payment from the company. METR also emphasizes that no compensation was conditioned on results and that there was no financial or organizational control by Anthropic over the evaluation. The displayed independence is therefore real, even if it does not dispel all reservations related to redactions and potential conflicts of interest.

Key points

  • METR publishes its report on November 5, 2026.
  • AECI shows an estimated acceleration of 1.9x since October 2025.
  • The confidence interval ranges from 1.5x to 2.8x.
  • Anthropic redacted most of the evidence deemed most useful.
  • The report is based on 11 weeks of investigation and about 15 interviews.
  • METR judges that AI probably explains most of the current acceleration.

By the numbers

  • 1.9x — estimated acceleration of benchmarks since October 2025, according to METR.
  • 95% — confidence interval published by METR: 1.5x to 2.8x.
  • 11 weeks — duration of the evaluation, from August 12 to October 29, 2026.
  • 2.5 FTE — size of the METR team mobilized for this work.
  • 15 interviews — approximate number of exchanges conducted with staff members.

To read