Skip to content
NewsResearch

METR finds Anthropic's AI models are speeding up its research about twice as fast as a year ago

· by

The short version

An 11-week independent assessment found Anthropic's benchmark scores rising 1.9 times faster than before October 2025, with AI itself behind most of the gain, though Anthropic redacted much of the underlying evidence.

METR finds Anthropic's AI models are speeding up its research about twice as fast as a year ago

An independent safety group that spent 11 weeks with privileged access inside Anthropic concluded the company's AI models are now improving on measurable benchmarks at close to double last year's pace, and pinned most of that speedup on AI itself rather than more training compute, per METR's November 5 report.

What

METR, an independent AI safety research organization, ran the assessment from August 12 to October 29, 2026 at Anthropic's invitation, reviewing internal benchmark data, training-compute metrics, and staff interviews the company does not normally disclose. Anthropic also shared forecasts of how future internal models were expected to score on those same metrics, a level of access METR called broader than initially scoped. Benchmark scores for Anthropic's models, measured on the firm's own capability index, grew at roughly 1.9 times the rate seen before October 2025, with a 95 percent confidence range of 1.5 to 2.8 times, which METR said amounts to "roughly two years of historical progress over the last year" (per the report). The group said it is confident that jump did not come from scaling up training compute. Instead, METR estimated, with substantial uncertainty, that about 80 percent of October's acceleration traces to AI use itself, a share it said has grown over time. Internal Slack polls cited in the report found Anthropic staff self-reporting being about 12 times as productive on their work using the company's most capable internal model as they would have been without AI assistance, roughly double what employees reported a few months earlier using an internal model called Mythos 5, according to The Verge's review of the findings. METR added it saw no sign of bottlenecks in compute, data, or algorithmic progress over the next 12 months that would slow the trend, raising the odds of faster acceleration still to come.

Anthropic redacted most of the evidence behind its own numbers

Anthropic had contractual rights to redact nonpublic information from METR's public report, and the group said the company "redacted a majority of the evidence we found most informative" for its risk conclusions. That leaves outside readers unable to check the mechanics behind the acceleration figure, only METR's summary judgment about it. Anthropic cofounder Sam McCandlish defended the decision in a post on X, writing that disclosing exactly what is driving the speedup "isn't the right thing to do" because the details are commercially sensitive and would hand rivals a roadmap to move faster themselves, as The Verge reported. For anyone trying to model how fast frontier labs are actually moving, that leaves a real gap: the topline multiplier is now public and independently vetted, but the evidence underneath it is not, so a reader has to take METR's word for the number rather than check the work.

Context and reactions

The report landed as Anthropic prepares to go public after filing an S-1 with the SEC, and only months after CEO Dario Amodei publicly urged labs to slow AI development. Three of Anthropic's newest models, labeled Model 2, Model 3, and Model 4 in the report, finished training in July, August through September, and September through October of 2026, and are already contributing to the R&D gains METR measured; the systems behind today's speedup have not shipped to the public. Marius Hobbhahn, CEO of Apollo Research, told The Verge the pace described in the report moves faster than the one envisioned by AI 2027, an influential forecasting scenario, even though the overall direction broadly matched what researchers in the field already expected. Alan Chan, a senior research fellow at the policy group GovAI, told The Verge the findings should push skeptics to shorten their timelines for when AI systems could fully automate tasks like software engineering. OpenAI said in September it is working toward an AI system that could independently advance its own deep-learning and alignment research by March 2028, which puts Anthropic's numbers inside a wider industry pattern rather than an isolated case. The report has also drawn criticism over an apparent tension in Anthropic's position: the company has campaigned publicly for cautious, coordinated AI development, yet METR's findings show it pushing the same technology toward thresholds that Amodei himself has warned about, per The Verge's coverage of the reaction.

What to watch next

McCandlish said Anthropic plans to publish its own report on the pace of AI progress "in the coming weeks," alongside new safety steps, which would give outside researchers a second data point to weigh against METR's figures. Also watch whether OpenAI or Google DeepMind grants a comparable outside group the same kind of access; METR's report covers only Anthropic.

Sources