METR says Anthropic's AI progress has sped up, with AI now helping build AI

What METR reported

METR is an independent AI safety research group. As technews.tw reported, on November 5 it released an evaluation saying Anthropic's models have improved about twice as fast over the past year on tasks that can be clearly defined.

The team puts the rate at roughly 1.9 times the earlier pace since October 2025. In plain terms, that is about two years of past progress packed into one.

AI helping to build AI

METR says the speed-up comes mainly from Anthropic using AI heavily in its own research and development, not simply from throwing more computing power at training.

The report also notes that more oversight and training work is being handed to other AI models. This is called recursive self-improvement: AI helping to improve the next AI, which can keep progress accelerating. METR has not yet seen a clear bottleneck that would slow this down.

What staff say inside the company

METR combined internal data, interviews and performance data from unreleased models over 11 weeks, with unusual internal access. Anthropic finished 3 models in the past half year that the public cannot use, and employees work with versions well ahead of what customers get.

In an internal Slack poll and interviews, staff reported big productivity gains. For the strongest internal model, the average self-rated boost was about 12 times. Some researchers said their bottleneck has moved from writing code to coming up with ideas and getting enough computing power.

The limits of what we can see

The report itself says Anthropic required large cuts, so outsiders get only part of the evidence and conclusions. The 12 times figure is also self-reported by employees, not a measured result.

METR adds that AI still cannot make major contributions on its own in the most important training steps. People remain in the loop where it counts most.