METR: Opus 5.5 is a step, not a leap
On 22 September 2026, the day Claude Opus 5.5 was released, METR published a summary of its pre-deployment evaluation: the model speeds up AI research slightly more than Claude Fable 5.1 but is unlikely to automate it fully, and on quantitative evaluations it is an incremental improvement, not a jump. A separate METR team with wider access to Anthropic estimated the acceleration of capabilities due to AI at about 1.5 times, with a roughly 30% chance of twice.
Why it matters
A quantity only the labs themselves had spoken of, how far AI already speeds up the making of the next AI, received an estimate from an outside organization. The estimate is preliminary: the period it refers to is not named, the authors of the summary did not see the evidence, and METR itself calls it a trial version of a broader assessment.
The evaluation ran under an unpaid agreement; Anthropic could review and edit the text, which also appears in the Opus 5.5 system card. API access lasted 10 business days, with five tasks: Budget NanoGPT Speedrun, LMCA, Train a Program, Gaming Bot and Sunlight. METR writes that the model keeps qualitative weaknesses an expert human is unlikely to show on hard long tasks, and that full automation of research will need large improvements in foresight and research judgement. Anthropic presents Opus 5.5 as performing at the level of Fable 5.1 on most work, costing 40% less to run than Opus 5, and as its first release since the call to pace the frontier. What the record does not claim: that METR's evaluation is fully independent of the developer, since the text was agreed with it, or that the figure of 1.5 applies to the development of Opus 5.5 specifically; METR itself says this is unclear.