Technology news and discussion on TalkFuse.
Anthropic’s models improving nearly twice as fast with help from AI, study finds
An independent study by AI safety group METR found Anthropic’s benchmark capabilities have been improving nearly twice as fast since October 2025. The findings suggest AI is helping speed up the development of more capable AI models, raising questions about whether safety assessments can keep pace.
METR estimated AI may account for about 80 per cent of the acceleration, but said that figure is uncertain and found no clear evidence of major bottlenecks over the next year. Its 11-week assessment used confidential research, interviews and data from unreleased models; it also found AI cannot yet make highly significant, fully autonomous contributions to the most important stages of model training.
In short
- Anthropic’s benchmark capabilities improved 1.9 times faster.
- METR estimates AI may explain about 80 per cent of the acceleration.
- Researchers say AI cannot yet independently drive key training stages.
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
If AI is helping researchers improve models faster, that may bring useful capabilities sooner and make research more productive. METR’s estimate is uncertain, and its study found no clear evidence of major bottlenecks in the coming year; it also found AI cannot yet make highly significant, fully autonomous contributions to the most important training stages. These limits support a measured response based on continued assessment rather than assuming rapid, unchecked acceleration.
The case against
Even partial AI assistance could shorten the time between model generations and make it harder for safety evaluations and oversight to keep pace. METR estimates AI may explain about 80 per cent of the acceleration, though the figure is uncertain, and its reliance on confidential material and unreleased models makes the findings difficult to independently scrutinise. Given the potential consequences of capability gains, advocates of precaution may argue for stronger evaluation and safeguards before development speeds further.
AI Research Science Technology World