METR finds Anthropic's model progress roughly doubled, mostly because of AI-assisted research
After eleven weeks inside Anthropic's internal data, the nonprofit says the company may have crossed its own acceleration warning threshold in July. Anthropic redacted key evidence.
The nonprofit METR says the rate at which Anthropic's models are improving has roughly doubled, and that about 80 percent of the speed-up comes from Anthropic using AI in its own research, The Decoder reports. From August 12 to October 29, 2026, a METR team of 2.5 full-time staff had what it calls unprecedented access to benchmark results for unreleased models, training metrics, employee surveys and internal analyses. METR's conclusion is that Anthropic may have crossed the acceleration warning threshold in its own Responsible Scaling Policy as early as July. That policy treats a doubling in the pace of AI progress caused by automated AI research as a risk indicator that should trigger a warning level. Researchers in the safety debate treat AI that speeds up the building of better AI as a key early warning sign. The worry is a feedback loop in which models improve their successors faster and faster. METR says such a loop could lead to a "sudden emergence of extreme and risky capabilities." According to The Decoder, Anthropic redacted the most important evidence in the report, which limits how far outsiders can check the findings. The assessment matters to regulators, investors and other labs because it is a rare outside check of a frontier developer against its own published safety commitments.