On November 5, 2026, METR published its review of the “Risks from automated R&D” section of Anthropic’s August 2026 Risk Report. METR agrees that catastrophic risk from models available by July 15, 2026 was low, but finds it plausible that those models had already crossed the Responsible Scaling Policy’s early-warning threshold of doubled AI R&D progress.
METR’s own estimates of leading indicators range from 1.3x to 2.8x, and a figure in Anthropic’s private report implies about 1.97x by METR’s calculation. METR says the public report’s “less than a factor of 2” implies a meaningfully lower estimate than the data supports and that “around 2” would be more accurate. The private report calls it plausible that counterfactual acceleration exceeded 2x, with a central estimate slightly under 2x, and METR broadly agrees with it rather than the public framing.
On Anthropic’s AI capabilities trend, the internal “Model 2” scores about 1.5 points above Claude Mythos 5; estimated acceleration was about 1.5x on the report’s data and about 1.9x after Anthropic added benchmarks. METR finds Anthropic’s argument that models could not fully replace its research staff broadly compelling, though with less confidence than before because benchmarks have saturated, and cannot rule out substantial further acceleration over the next six months.
Why it matters: it is an external evaluator, with access to non-public material, publicly disagreeing with a lab’s headline reading of its own risk threshold - the kind of check third-party review of risk reports is meant to provide.
What it does not show: Anthropic redacted non-public information, including METR’s own leading-indicator analysis; METR stresses that counterfactual acceleration is very hard to measure; the review’s scope was negotiated during the engagement; and Model 2 had not finished training.