METR Study Finds Anthropic's Models Are Improving Nearly Twice as Fast, With AI Now Helping Build AI

Anthropic's models are getting better at roughly twice the clip they used to, and a big part of the reason is that AI is now helping build the next generation of AI. That's the headline finding from a new independent study by METR (Model Evaluation and Threat Research), an AI safety nonprofit that's become one of the more credible outside auditors of frontier labs.

The study ran for 11 weeks, from August to October, and METR got access most outsiders never see: confidential research, employee interviews, and performance data from unreleased models. The conclusion is blunt. Anthropic's capability improvements on well-specified benchmark tasks have been running about 1.9 times faster since October 2025 than they did in the prior stretch, from March 2024 through October 2025.

"As of October 2026, Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025," METR wrote. The group was careful to flag the uncertainty baked into that number: "We would not be surprised if the true acceleration was closer to 1.5 or 2.5 the historical rate."

Here's the part that matters most for anyone tracking the trajectory of frontier AI: METR explicitly ruled out the usual suspect. Compute scaling isn't driving this. "We are confident that the acceleration was not due to an acceleration in the scaling of training compute," the group said. Instead, the gains appear tied to AI systems increasingly participating in the research and engineering loop that produces better AI systems.

How AI Is Speeding Up Its Own Development

This is the mechanism researchers have been calling recursive self-improvement, or RSI for short. The basic idea: once AI gets good enough at tasks like writing code, running experiments, and analyzing results, it can start contributing meaningfully to the process of building its own successors. Do that well enough, and you get a feedback loop where each generation of models helps produce the next one faster than humans alone could manage.

Anthropic's own researchers have reported exactly this kind of productivity boost internally. Engineers and researchers are reportedly leaning on AI tools to write code and run experiments, and informal employee polls pointed to productivity improvements of around 12 times. METR was quick to caution that number shouldn't be taken at face value. Self-reported estimates like that are notoriously squishy, shaped by selection bias, vague baselines, and the natural tendency to round up anecdotes into tidy multiples.

A few things are worth unpacking for anyone trying to gauge how far along this loop actually is:

• METR found no evidence that AI is yet capable of making highly significant, fully autonomous contributions to the most important and difficult stages of model training. In other words, humans are still firmly in the loop for the hard parts.
• The 1.9x acceleration figure is a benchmark metric, specifically about performance on well-specified tasks. That's narrower than "AI is twice as smart" or "AI is twice as capable across the board."
• The timeframe matters too. This acceleration is measured against Anthropic's own historical baseline, not against competitors, so it says more about internal trajectory than industry-wide pace.

Taken together, that paints a picture that's significant but not yet alarmist: AI is clearly accelerating its own development pipeline, but it hasn't crossed into the territory of fully autonomous self-improvement that worries safety researchers most. The gap between "helping a lot" and "doing it without us" is still there. It's just narrower than it was a year ago.

That gap is exactly why this finding is landing the way it is. Anthropic CEO Dario Amodei called for slowing down advanced AI development back in September, a notable stance from someone running one of the labs pushing hardest at the frontier. Findings like METR's give that call some empirical weight. If capability gains are compounding faster than expected, and the mechanism driving that compounding involves less human oversight at each step, the argument for caution gets harder to wave off as hypothetical.

The stakes here extend well beyond one company's benchmark curve. AI researchers have spent recent months warning that systems capable of meaningfully accelerating their own development could become progressively harder to align with human goals, monitor effectively, and control if things go sideways. That's not a new concern, but concrete evidence of acceleration, measured by an outside party with real access, is different from theoretical hand-wringing about future risk.

There's also an industry-scale angle worth flagging. The AI sector has poured hundreds of billions of dollars into compute infrastructure on the premise that scale is the primary lever for progress. If AI-assisted R&D becomes a second major driver of capability gains, independent of raw compute scaling, that could reshape how labs think about where to invest and how fast the field moves overall.

All of this comes as Anthropic prepares for a public listing, potentially as soon as this month, in what could be one of the largest IPOs in tech history. Investors sizing up that offering now have a genuinely interesting data point: independent, outside verification that the company's core product, model capability itself, is improving at an accelerating rate. Anthropic did not respond to a request for comment on the study.

Key Takeaways

METR's study is one of the clearest outside-verified signals yet that frontier AI development is starting to feed on itself, with models contributing meaningfully to building their successors rather than just being products of human-directed research. The 1.9x acceleration is real but bounded. It's not driven by compute scaling, and full autonomous self-improvement hasn't arrived. Still, the trendline is the story here: if the pace of self-assisted development keeps compounding, the window for external oversight to keep pace may keep shrinking too.