What is AI R&D acceleration? METR estimates Anthropic's pace of progress has about doubled.

Anthropic is likely improving its AI models on well-specified tasks at about twice the rate it did between March 2024 and October 2025, according to an assessment released Thursday by METR, a nonprofit that evaluates AI systems.

METR called the estimate highly uncertain. It said it "would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate."

The report comes as Anthropic plans to go public later this month. In September, CEO Dario Amodei called for a slower pace, writing, "We must slow the pace at which we improve the capabilities of AI models."

What does acceleration mean?

METR defines a 2x acceleration as "2 years of historical progress" being made in one year. The figure measures how quickly models are improving; it doesn't, on its own, say why. More computing power, more staff, or better AI tools could all contribute.

The related term is recursive self-improvement, which Amodei has described as "AI's growing ability to build the next generation of AI." He wrote, "Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all."

METR tracked benchmark scores

METR reviewed Anthropic's internal data and interviewed staff from Aug. 12 to Oct. 29. Its main measure was AECI, Anthropic's version of the Epoch Capabilities Index, which turns benchmark test results into one score. METR estimated that the trend rose from 15 points a year before October 2025 to 29 points a year after it.

Source: METR and Anthropic. METR said its figure tops the Haiku 5.5 card's mainly because it includes newer unreleased models.

The confidence interval captures uncertainty in model scores, not every uncertainty in the assessment. METR's broader judgment allows for values closer to 1.3x or 3x.

Much of METR's evidence came from tasks "easy to grade automatically or with a language model." If models improved faster on those clean tasks than on messy ones, METR said, "we may be overstating the acceleration in real-world impact." METR also said Anthropic's models grew more slowly than the industry frontier before late 2025, which could explain some of the speedup.

METR attributes most of the speedup to AI

METR's subjective estimate is that a majority of the acceleration as of October 2026, perhaps about 80%, is due to AI. That share applies to the increase in Anthropic's growth rate at that point, not to all of its progress. METR said it generally expects AI's cumulative share since October 2025 to be lower, partly because non-AI factors likely played a large role in late 2025.

METR said the 80% estimate carries substantial uncertainty. It also said it was confident the acceleration didn't come from faster scaling of training compute.

Anthropic staff voted in Slack polls on their estimated speedup from the most recent model, and the average reached about 12x in October. METR defines speedup as the ratio of a staff member's or team's output with AI to their output without it, though the report doesn't give the poll's wording. Speedup is a different measure from Anthropic's overall rate of progress, and METR said individual speedup estimates don't correspond to acceleration rates. It called the polls "only very loose evidence."

Interviews with researchers produced a wide range of estimates centered around 2x, METR said.

METR said, "Fully autonomous contributions to AI R&D are not currently a large contributor to overall progress, but they may be in the near future." It didn't identify any convincing bottlenecks, but said it isn't confident in that assessment.

Anthropic controlled what METR could publish

METR said it took no payment from Anthropic, which provided laptops and free tokens. Anthropic had unilateral redaction rights over non-public information, and METR said the company "redacted a majority of the evidence we found most informative for these conclusions." Anthropic also dropped a planned check of its compute allocation midway through the project.

This project is separate from the embedded evaluator arrangement we covered in October, under which METR can publish without Anthropic's approval. Anthropic co-founder Sam McCandlish wrote on X that sharing what's driving the speedup is "commercially sensitive" and "risks handing out a roadmap for others to go faster."

Anthropic's Sept. 22 system card for Claude Opus 5.5, the source of the 1.53x in the table, said its internal measures didn't show a sustained twofold acceleration attributable to AI. METR's headline figure covers total acceleration, so the two aren't the same measure. In a separate review, METR said Anthropic's August public risk report "overall understates the level of acceleration that was occurring," while agreeing that catastrophic risks from its models as of July 15 were low.

McCandlish wrote, "AI is speeding up our research, but it's hard to say exactly how much." He said Anthropic plans to release its own analysis in the coming weeks.