Anthropic is improving the capabilities of its AI models roughly twice as fast as it was a year ago, and much of that speedup appears to be coming from AI systems helping to build their own successors, according to an independent analysis released Thursday by the nonprofit evaluation group METR.
The findings, reported by the Associated Press, carry unusual weight because they rest on what METR described as "unprecedented access" to Anthropic's internal data. Researchers from the Berkeley, California-based group — whose full name is Model Evaluation and Threat Research — were invited inside the company for roughly 11 weeks, from mid-August through late October, where they analyzed internal research data and interviewed staff.
For readers tracking the safety debate around frontier labs, see our breaking AI news coverage.
The Core Finding: Progress Is Compounding
METR's central estimate is that Anthropic is improving its models' performance on well-specified tasks about twice as fast as it was a year earlier. The report put it starkly: "Anthropic appears to have made as much progress on benchmarks in the past year as they did between October 2023 and October 2025."
More consequential is what METR believes is driving the acceleration. The researchers said they were able to "confidently" rule out growth in computing power as the cause — traditionally the variable most tightly linked to model performance. Their conclusion: the improvement is largely down to using AI itself to do AI research.
Anthropic has acknowledged that its Claude chatbot increasingly helps the company develop newer, more advanced versions of itself. METR's report found no "bottlenecks" — such as restrictions on compute or data — holding that process back, and noted that while AI is not yet making "highly significant and fully autonomous contributions" to the most important parts of model training, Anthropic researchers described significant "signs of life" implying that could change soon.
The Findings Landed Mid-Slowdown Debate
The timing is awkward for a company whose chief executive has been lobbying for restraint. It was during METR's assessment period that Anthropic CEO Dario Amodei published a September 12 essay urging the AI industry to "slow the pace at which we improve the capabilities of AI models" so that safety measures could catch up.
METR also reviewed both the public risk report Anthropic published in August and a private version, and concluded that the public document "overall understates the level of acceleration that was occurring."
Anthropic's Response: Partial Agreement, Heavy Redactions
Sam McCandlish, an Anthropic co-founder and the company's chief architect, responded to the report Thursday in a lengthy post on X, writing that many of METR's conclusions match his own view.
"AI is speeding up our research, but it's hard to say exactly how much," he wrote, adding that Anthropic plans to release its own analysis of the pace of AI progress inside the company in the coming weeks.
McCandlish also defended the company's decision to redact large portions of the report, which limited how precisely METR could communicate its acceleration estimates. He said sharing "exactly what is driving the speedup isn't the right thing to do," because the information is both commercially sensitive and "risks handing out a roadmap for others to go faster" — which, he argued, would make the very risks METR is worried about worse. "That's a tradeoff," he wrote.
Important Caveats
The report was, by METR's own description, peppered with caveats. The group said its headline finding could be overstating matters: the measured acceleration may apply only to well-specified tasks that can be graded automatically, rather than the messier, open-ended work that defines frontier research. Alternatively, Anthropic may simply have understated its earlier baseline, making the latest jump look bigger than it is.
METR's position is also worth noting for transparency reasons: the nonprofit is funded through donations, does not accept compensation for its evaluations, and has conducted capability assessments for Anthropic, OpenAI and other labs for years — though partner companies do provide model access and the tokens needed to run its research.
Why It Matters Now
The analysis lands weeks before Anthropic expects to go public in what could be the largest initial public offering in history, with debate swirling over whether increasingly capable AI models might elude human control. Safety advocates have long argued that independent verification — not lab self-reporting — is the only trustworthy measure of how fast the frontier is moving, particularly as the Trump administration has encouraged companies to police themselves.
METR flagged one final concern: if AI-assisted AI research begins to compound, acceleration itself could accelerate, with the risks from advanced models growing at a pace outside observers cannot easily track. The group's report suggests that for one leading lab at least, the feedback loop the field has worried about is no longer hypothetical — it is showing up in the benchmark data.
How quickly the rest of the industry follows is the question regulators, investors and the company's own executives will now be watching.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
