Anthropic’s AI gains on well-defined tasks are accelerating, METR estimates

Anthropic’s models are getting better at well-defined tasks about twice as quickly as a year earlier, according to a METR analysis released November 5. The finding puts a measure on accelerating progress inside a company whose chief executive has urged the industry to slow down—and raises questions about how much AI itself is driving that acceleration.

The assessment period overlapped with Dario Amodei’s September 12 essay calling for slower capability gains so safety work could catch up. He also backed embedding outside evaluators in leading labs and singled out METR as a suitable organization. The new findings therefore concern both the pace he warned about and the kind of outside scrutiny he advocated.

Eleven weeks inside the research process

Anthropic invited Model Evaluation and Threat Research, a Berkeley-based nonprofit, to examine internal data and interview employees. One of its reports drew on company-provided data covering about 11 weeks, from mid-August through late October. METR described the access to Anthropic’s research and development process as unprecedented.

The researchers asked two questions: whether the growth rate of Anthropic’s internal capabilities had accelerated, and how much of any acceleration came from AI. Their headline estimate concerns progress on well-specified tasks, not a blanket doubling of what every model can do or how quickly the company releases products.

Anthropic appears to have made as much progress on benchmarks in the past year as they did between October 2023 and October 2025.

METR report, quoted by the Associated Press

Why AI assistance is the leading explanation

METR said it could confidently rule out growth in the computing power used to train models as the reason for the acceleration. Its researchers instead suspected that AI was helping speed up AI development. Anthropic has said Claude increasingly helps the company develop newer, more advanced versions of itself.

That explanation remains a hypothesis, not a demonstrated attribution of all the gains. METR also drew a boundary around the autonomy involved: AI was not yet making highly significant, fully autonomous contributions to the most important parts of model training. Anthropic researchers nevertheless described early signs suggesting such contributions could arrive soon.

The concern is a compounding cycle: better AI helps build better AI, which then speeds up the next round of development. METR warned about that possibility and said it found no bottlenecks, such as computing-power or data restrictions, holding back Anthropic’s progress. Amodei’s September essay similarly warned that AI’s growing ability to build its successors could outrun human understanding and control.

The estimate has boundaries—and disclosure limits

METR cautioned that its estimate could overstate the acceleration. It identified two ways the measured gains might give an incomplete picture:

• The acceleration might be confined to tasks that can be clearly specified and graded automatically or by a language model, rather than messier work.
• Anthropic might have understated its earlier progress, making the more recent improvement look larger by comparison.

There were also internal disagreements and uncertainty within METR’s team. Anthropic’s redactions limited how much detail the nonprofit could publish about its acceleration estimates, according to the Associated Press. Anthropic did not immediately respond to the news agency’s requests for comment.

METR is funded by donations and does not accept compensation for its evaluation work. Its relationships with AI labs still provide essential resources: partner companies have supplied access and tokens used for evaluations, research and engineering.