01 / AI · Claim vs. Check · Analysis
METR estimates 1.9x faster benchmark progress at Anthropic
METR attributed about 80% of the current acceleration to AI use, but Anthropic redacted a majority of the evidence behind that judgement.

METR estimated Anthropic's AECI growth at 1.9 times its pre-October 2025 rate in an assessment published on 5 Nov 2026, with a 95% confidence interval of 1.5 to 2.8 times. AECI is a benchmark index, Anthropic's own variant of the Epoch Capabilities Index.
The report used internal company data to measure how fast AI is speeding up AI development. METR guessed that about 80% of the acceleration comes from AI, but Anthropic redacted most of the evidence METR found most useful for its judgement.
METR judged the size of the acceleration, its cause and the risk over the next year in turn. Each judgement carries a caveat about uncertainty.
Anthropic AECI acceleration estimates compared
| METR estimate | 1.9 | |
|---|---|---|
| Haiku 5.5 system card | 1.7 |
The estimate of two years' progress made in one
METR staff reviewed Anthropic's internal data and interviewed relevant employees from 12 Aug 2026 to 29 Oct 2026. Most of the staff involved judged that capabilities on well-specified tasks were improving at around twice the pace of March 2024 to October 2025. The report said it would not be surprised if the true value was closer to 1.3 times or 3 times.
In the report, a 2x acceleration means making two years of past progress in one year. On that estimate, Anthropic made as much benchmark progress in the past year as it did between October 2023 and October 2025. In model terms, that span runs roughly from Claude 2.1 to Sonnet 4.5.
METR compared its estimate with 1.7 times, the company-published figure in the Haiku 5.5 system card. It attributed the gap mainly to the inclusion of data from new, unreleased models. METR's reanalysis, which varied conditions such as benchmark and model selection, ranged from 1.1 to 2.9 times.
The report noted that the 1.9 times figure could reflect benchmark overfitting or be an artefact of the method. The evidence that led METR to take a 2x acceleration more seriously was redacted. METR also wrote that further uncertainty surrounds how far the acceleration carries over to broader, less well-defined tasks.
A guess that AI drives about 80% of the acceleration
METR guessed that about 80% of the acceleration as of October 2026 came from AI use, and added that the uncertainty is substantial. It wrote that non-AI factors likely influenced the acceleration in late 2025. Its detailed estimates of cumulative and instantaneous effects were redacted.
METR stated that faster growth in training compute, the computing resources used to train models, was not the cause. The detailed evidence was redacted. Anthropic also redacted detailed estimates of AI's effect at each training stage, so they are absent from the public document.
METR classed staff self-reports of speed-up as very loose evidence. Slack poll figures reached 12 times in October 2026, but they record the speed-up individuals perceived and are a different value from the acceleration rate. METR's analysis of how much individual speed-up would be needed was redacted.
METR was told that R&D acceleration came through AI assistance more than through fully autonomous contributions. It also wrote that there are signs fully autonomous contributions will soon grow. The company has not disclosed when.
A risk judgement that rests on redacted evidence
In an October 2026 article, METR set out questions the public needs to understand about AI R&D acceleration. This report records how its judgement on each question changed. METR's stated reason is that a large AI-driven speed-up in AI development could cause dangerous capabilities to emerge suddenly.
METR wrote that it found no major bottleneck, a limiting factor that holds back growth, that would take effect within the next 12 months. Its view of bottlenecks changed several times during the assessment, and it is not confident in this finding. With no clear bottleneck identified, METR judged that there is a significant risk of acceleration growing faster over the coming year.
METR was free to publish its top-level conclusions about risk. Anthropic redacted most of the evidence METR found most useful, and described the redactions as removing confidential information that could speed up other AI developers. METR wrote that some redactions are themselves hard to describe.
Anthropic also redacted the analysis of individual bottlenecks, so it is absent from the public document. The public cannot readily examine this evidence for itself. METR wrote that the report does not achieve transparency about evidence, and that the public must rely heavily on METR's judgement.