Skip to content

Research Story

METR estimates Anthropic has probably roughly doubled its pace of model improvement on well-specified tasks

METR, in an assessment published on 5 November 2026, estimates that Anthropic probably improved model capabilities on well-specified tasks at about twice the March 2024 pace through October 2025, and that AECI scores have grown at 1.9 times the earlier pace since October 2025.

Published

This story was written with the help of artificial intelligence, checked automatically, then reviewed by an editor before publication, under our AI policy.

What METR estimates

In an assessment published on 5 November 2026, METR estimates that Anthropic has probably improved the capabilities of its models on well-specified tasks at about twice the rate seen in March 2024, across the period from March 2024 to October 2025. METR describes the uncertainty as strong, with a plausible range between 1.3x and 3x.

A separate estimate covers the period since October 2025: METR puts the growth in AECI scores at 1.9 times the earlier pace, with a 95% confidence interval of 1.5 to 2.8. The two estimates cover different periods and different measures (well-specified tasks for the first, AECI scores for the second).

These figures are METR's own estimates, not measurements published by Anthropic. The facts available do not say how far they carry over to other kinds of research work.

How much of the acceleration comes from AI

METR estimates that about 80% of the current acceleration, as of October 2026, is due to AI, again with substantial uncertainty. According to METR's assessment, most of the current speed-up is therefore attributed to AI.

METR also reports that it found no clear bottleneck and no strong diminishing returns that would prevent a faster acceleration. It acknowledges uncertainty on this point, so the finding is an absence of evidence for such limits rather than proof that none exist.

Access and redactions: the limits of the assessment

METR evaluated Anthropic from 12 August to 29 October 2026 and had extended access to internal data. Anthropic had the right to write caveats and to censor non-public information. In METR's words: “Anthropic had rights to redact any non-public information in this post.”

According to METR, Anthropic redacted a majority of the evidence METR judged most informative. Anthropic justifies the redactions as “aimed at removing confidential information that could accelerate other AI developers.”

This matters for reading the results. Readers of the public post cannot inspect much of the strongest evidence behind the estimates, and the facts available do not say which items were removed or how METR's conclusions would differ without the redactions. Independent verification of the figures is therefore limited.

What remains open

Several questions are not answered in the material available: how AECI scores are constructed, what the well-specified tasks cover, and whether the acceleration will continue. The width of the intervals (1.3x to 3x for the earlier period, 1.5 to 2.8 for the recent one) shows that the central figures are approximations.

The primary source is METR's post at metr.org.

METR's estimates of Anthropic's acceleration in capability improvement (data as of )
PeriodMeasureEstimateUncertainty
March 2024 to October 2025Capabilities on well-specified tasksAbout 2x the March 2024 pace1.3x to 3x
Since October 2025AECI scores1.9x the earlier pace95% CI: 1.5 to 2.8

FAQ

How much faster is Anthropic improving its models, according to METR?

METR estimates about twice the March 2024 pace on well-specified tasks from March 2024 to October 2025 (range 1.3x to 3x), and 1.9 times the earlier pace for AECI scores since October 2025.

How much of the acceleration is attributed to AI?

METR estimates that about 80% of the current acceleration, as of October 2026, is due to AI, with substantial uncertainty.

Was the assessment independent of Anthropic's editing?

Only partly: Anthropic could write caveats and redact non-public information, and according to METR it redacted a majority of the evidence METR found most informative.

Sources

  1. https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/, metr.org, (primary source)

Plain text version (Markdown)