METR finds Anthropic's AI research is now moving twice as fast as last year

This site reported earlier that Anthropic had released FlashProve, a prototype for proving AI responses. A new report published this week turns to a different question: how much Anthropic’s AI systems now drive the pace of Anthropic’s research.

The AI safety organization METR published its findings on November 5, after an 11-week investigation inside Anthropic evaluating its research and development. METR found Anthropic’s models are improving roughly twice as fast as a year ago, with benchmark scores suggesting about two years of historical progress packed into the past twelve months.

METR said it is confident the acceleration is not simply the result of training with more computing power, which has historically driven AI progress. Instead, the researchers said most of October’s speedup came from AI itself, with AI’s share growing over time.

That conclusion was backed by Slack polls and staff interviews. Employees using Anthropic’s most capable internal model reported being, on average, around 12 times as productive as without AI help, roughly double what they reported a few months earlier using a model called Mythos 5. METR added it found no clear bottleneck likely to slow the acceleration.

Much of that gain comes from internal models the public has not seen. METR said Anthropic finished training three unreleased systems in the past six months, code-named Model 2, Model 3 and Model 4, completed in July, August-to-September and September-to-October, plus an older unreleased model from 2022.

METR said Anthropic redacted most of the evidence it found most informative from the public version of its report, citing commercial sensitivity. Anthropic cofounder Sam McCandlish said on X that this makes outside checks harder, but that disclosing the exact drivers would hand rivals a roadmap and worsen the risks METR raised. He said Anthropic plans its own report on AI’s pace of progress, plus new safety steps, in the coming weeks.

The findings drew sharp reaction from safety researchers. Marius Hobbhahn of Apollo Research told The Verge the pace is faster than forecasts in AI 2027, a widely discussed scenario for AI beginning to improve itself with little human input. Alan Chan of the Centre for the Governance of AI said the findings should push skeptics to shorten their timelines for AI automating software engineering. METR researcher Ryan Greenblatt wrote on X that he had underestimated how fast such self-improvement would arrive in 2026.

What this means

The report lands as Anthropic prepares to go public and as its chief executive, Dario Amodei, has publicly urged AI labs to slow down, a contrast the findings sharpen rather than resolve. Inside the company, most AI-assisted research still means AI helping human researchers rather than working alone, though METR warned fully autonomous contributions could soon play a larger part.

What remains unclear is exactly what is producing the speedup, since Anthropic’s redactions removed much of the underlying evidence, and METR said that leaves the public relying on its judgment rather than checking the findings directly. The reports do not say whether Anthropic’s promised follow-up report will restore any of the material METR says was cut.

OpenAI said in September it is making progress toward an AI system able to advance its own research by March 2028, a sign Anthropic is not alone in pursuing this kind of self-reinforcing development.