METR estimates Anthropic’s AI progress nearly doubled over the past year

On November 5, 2026, METR said Anthropic’s benchmark scores were improving at 1.9 times their earlier rate, measured since October 2025 — the central finding of the group’s 11-week investigation into how the AI company builds its models. In METR’s framing, Anthropic packed roughly two years of the older pace of progress into the past year.

On the benchmark acceleration, the group’s 95 percent confidence interval spans 1.5 times at the low end and 2.8 times at the high end, and separate robustness checks spanned 1.1 times to 2.9 times with the result hinging on one modeling choice. Most of the group’s staff see Anthropic improving at roughly twice the historical rate, but the team said it would not be surprised by anything from 1.3x to 3x. According to METR, Anthropic’s trend broke on Oct. 7, 2025.

Before October 2025, Anthropic’s AECI benchmark scores rose by roughly 15 points a year; after that, the annual gain was roughly 29 points. On Model 4, AECI reached 178.6, and the model topped 50 percent on all 15 benchmarks METR reviewed, although the group said Model 4 lacks enough benchmarks, leaving it nervous about measuring AECI now and in the future.

METR estimated that Anthropic’s capabilities on specified tasks are improving at roughly twice the March 2024-to-October 2025 pace. AI use, METR judged, was likely responsible for most of the current acceleration, while training compute growth did not drive it. “We are confident that the acceleration was not due to an acceleration in the scaling of training compute,” METR said.

The group put AI’s share of the October 2026 acceleration at perhaps 80 percent, a judgment made with uncertainty. Anthropic’s own researchers described early signs of fully autonomous AI contributions to training arriving soon.

Recursive AI development means using today’s models to help build the next generation: writing code, designing experiments, and analyzing results so each version improves the one after it. Dario Amodei, Anthropic’s chief executive, described recursive self-improvement as the ability of AI to build the next generation of AI systems, and warned that without checks it could go beyond what humans can understand and control. TechBooky’s coverage framed the stakes this way: better AI could make the builders of the next generation more productive, shortening the development cycle and adding pressure on testing, external review, and governance.

Anthropic’s own documents point to a smaller effect. Its Claude Opus 5.5 system card put the slope increase at roughly 1.53x and said the slope had not doubled. The Haiku 5.5 system card put the figure at 1.7x. The private-report figure, per METR, was about 1.97 times, or around 2 times.

Anthropic’s Responsible Scaling Policy version 3.4, which took effect July 8, sets a 2x Acceleration Threshold, with the trigger at progress roughly doubling through AI research automation. The August 2026 Risk Report, which covers the period through July 15, said acceleration remained below a factor of two.

METR called overstated Anthropic’s case that the threshold had not been crossed, and said that by July 15, crossing the mark was plausible. “the argument that the threshold was not crossed is overstated,” METR said. Researchers at both METR and Anthropic mostly lean toward the view that the threshold was not crossed. METR warned that risk could be pushed high quickly if more acceleration arrives within six months, and the group also sees significantly faster acceleration over the coming year as a significant risk.

Anthropic used its unilateral redaction rights to take out most of the evidence METR considered most informative. The hidden material covered pipeline-stage acceleration estimates, forecasts for future models, compute-data-algorithm breakdowns, and bottleneck analysis. METR said readers cannot verify the evidence from the public version themselves, so they must rely on the group’s judgment.

According to Anthropic, the redactions aimed to strip out confidential information that might help rival AI developers move faster, and the company signed off on all public-report information without controlling the conclusions. Every information request from the group needed Anthropic’s unilateral approval or rejection, even as Anthropic gave the group access to its internal benchmarks, its training data on the research process, and staff interviews. The company also dropped a contract component meant to verify how it splits compute between capabilities and safety work.

METR said its view of constraints changed several times after Anthropic shared more detail, so limited access may have caused misread context. One person on the METR team received and turned down an Anthropic job offer, so an unconflicted colleague reviewed that person’s work. According to METR, Claude had a major role in collecting and analyzing data, though the group calls that only a possible source of distortion, and an unlikely one. Internal Slack polls at Anthropic pointed to an average self-reported speedup of about 12 times from its most capable model in October, and METR dismissed the poll results as weak and said the sample was probably unrepresentative.

From August 12 through October 29, the group ran its high-access evaluation at Anthropic. Looking at algorithmic progress, training compute, experiment compute, and data, the group saw no significant bottlenecks within 12 months, a conclusion it drew with low confidence.

METR started as an independent nonprofit in December 2023 under Beth Barnes, after splitting out of the Alignment Research Center team. Donations fund it, and it takes no payment for its evaluations. Over several years the group has assessed model capabilities at Anthropic, OpenAI, and other labs. It accepted no payment from Anthropic for the assessment, even though Anthropic provided free tokens and its staff used company devices.

The assessment is distinct from an October 15 arrangement under which six people on the METR staff received Anthropic hardware, office space, and internal tools. Under that arrangement, Anthropic cannot block publication, but it can demand redactions in set categories, among them confidential intellectual property. Anthropic said it had signed the White House Accord on Super Intelligence, under which companies agree to collaborate with independent external auditors or evaluators.

In September, Dario Amodei said Anthropic and its competitors should slow the improvement of AI capabilities so safeguards can keep up, and in a September 12 essay he called on the industry to reduce how quickly AI capabilities are improved. “slow the pace at which we improve the capabilities of AI models,” Amodei said. Amodei has championed embedding external evaluators in leading AI labs, and he specifically named METR as a fitting outside evaluator. OpenAI’s Sam Altman also supported giving independent evaluators employee-like access.

The company declined to comment on the report. The report’s conclusions largely match his own assessment, said Sam McCandlish, Anthropic co-founder and chief architect. The company plans to publish its own analysis of its AI progress pace in the coming weeks. The findings rest on one report: METR’s assessment of Anthropic after an 11-week investigation.

A planned Anthropic stock offering later this month could become the largest IPO ever. In an amended IPO filing dated Nov. 2, Anthropic said future models may soon meet the acceleration threshold, and said that no current model is designated as exceeding the mark. Its original S-1 filing came on Oct. 26. Tech Times reported that investors doubt the company can start its roadshow next week, and that finishing a November listing looks doubtful.