METR Finds AI Has Roughly Doubled the Pace of Anthropic's Model Progress

METR said in a report published on November 5 that growth on its capability measure rose from about 15 points a year to about 29, which it described as roughly two years of historical progress compressed into one. The 95% confidence interval runs from 1.5 to 2.8 times the pre-October 2025 pace. Its best guess is that about 80% of the acceleration seen by October 2026 is due to AI, though it stressed that the attribution is highly uncertain.

The finding goes to one of the risks AI labs and safety researchers watch most closely. If AI systems materially speed up the work of building better AI, capabilities could rise faster and more abruptly than companies, regulators or customers expect. Anthropic lists AI R&D acceleration as a key risk indicator in its Responsible Scaling Policy, although METR said it did not assess the results against that policy's thresholds and aimed for an overall picture instead.

Assistance Rather Than Autonomy

METR staff ran the engagement from August 12 to October 29, using Anthropic's internal benchmark scores, including scores for models that have never been released, along with algorithmic efficiency data, headcount and code volume figures, internal surveys and about 15 staff interviews. Capabilities were measured with AECI, Anthropic's version of Epoch AI's Capabilities Index.

Anthropic's own system card for Haiku 5.5 had put the acceleration at 1.7 times. METR said the gap comes mainly from its inclusion of internal-only models. Different statistical choices produced estimates ranging from 1.1 to 2.9 times, and measuring all frontier developers rather than Anthropic alone gave 1.4 times. METR also wrote that Anthropic's growth from 2023 to 2025 was low relative to the frontier, which inflates the ratio.

The evaluator ruled out one obvious explanation. "We are confident that the acceleration was not due to an acceleration in the scaling of training compute," the report said.

Staff told METR that the gains come mostly from AI helping researchers rather than from AI doing research on its own. "Fully autonomous contributions to AI R&D are not currently a large contributor to overall progress, but they may be in the near future," the report said, adding that some researchers described signs of life for autonomy. In informal Slack polls, Anthropic staff estimated their own uplift at about 1.2 times with Claude 3.7 Sonnet in January 2025 and about 12 times with an internal model METR calls Model 4 in October 2026. METR called this "only very loose evidence".

METR said it found no convincing bottlenecks in algorithms, compute or data that would slow progress over the next 12 months, though its view changed several times during the work.

What Anthropic Kept Back

Anthropic held unilateral redaction rights and removed most of the evidence METR considered most informative, saying the cuts were "aimed at removing confidential information that could accelerate other AI developers." As a result, METR wrote, "the public has to rely on METR's judgment much more than would be ideal." It did not have query access to unreleased models, and some of its requests were refused.

The company also narrowed the project partway through. The original contract included verifying how Anthropic splits compute between capabilities and safety work; Anthropic dropped that to keep the engagement focused on risk assessment and said it would revisit the question later. METR reached no conclusions on compute use.

METR disclosed that one team member had been in job talks with Anthropic and held equity in a competitor, and said that person's work was checked independently. It said it was not paid for the assessment but received free tokens and laptops.

METR judged it "extremely unlikely" that Anthropic deliberately misled it. The full report, about 100 pages, remains private, and the AI R&D section of Anthropic's August 2026 Risk Report is due to be published separately.