METR: Anthropic's AI models advance almost twice as quickly with AI assistance

what happened

An independent study by the AI safety nonprofit METR, published on Thursday, found that Anthropic's models have been getting better nearly twice as fast as before, as AI systems take a larger part in building their successors. The researchers estimate AI accounts for perhaps 80% of this speed-up, though they say the figure is uncertain. Reuters reported the findings.

METR's 11-week review drew on closed research, interviews and performance data from models not yet released. It concluded that benchmark gains have run 1.9 times faster since October 2025 than before. The authors saw no convincing sign of major bottlenecks that could slow progress over the coming year, and stressed that the speed-up was not caused by faster growth in training compute.

Still, AI cannot yet make truly consequential, fully independent contributions at the core stages of training models. Anthropic has reported strong productivity gains from using AI for coding and experiments, but METR considers employee surveys showing gains of roughly 12 times unreliable. The company is preparing to go public, potentially as soon as this month.

why it matters

For markets, the report adds fuel to the debate over whether huge spending on AI infrastructure is justified, ahead of what could be one of the largest IPOs ever. Faster capability gains heighten investor interest in the sector, but they also sharpen concerns about risks that could draw regulatory attention. The report does not directly point to price moves in any individual listed asset.

market impact

AI stocks and Anthropic's IPO: the data boosts sector interest but raises questions about costs and risks

tags

instruments