Anthropic flags possible AI research acceleration Anthropic flags possible AI research acceleration

A forecast, not a designation

The estimate was absent from the October 26 filing. Anthropic is still assessing the data. It also says it does not anticipate safely using Claude to build successive frontier models fully autonomously, without meaningful human supervision or understanding, if that becomes technically possible. It is considering policy changes to retain human control. The amendment also acknowledges that past acceleration became apparent only afterward.

What the threshold actually measures

The July policy defines two routes to its automated research threshold. One requires models able to replace all Anthropic research scientists and engineers at costs within a factor of five. The other is an observed or expected doubling of aggregate capability progress. The doubling is relative to both the expected rate and the fastest extended pre-automation rate, the latter spanning at least three model generations. Automation must plausibly explain a substantial share of the acceleration, rather than extra staff, compute or general productivity. Continued progress must also appear likely to accelerate further. The policy explicitly distinguishes this from doubling an individual researcher’s productivity. It separates Anthropic’s own mitigation plans from broader industry recommendations. Its company plans include centralized development logs, alignment assessments and security research. Full external review is required when a risk report covers a model assessed above the automated-research threshold and is significantly redacted.

The earlier measurement had limits

Anthropic’s August risk report covered activity through July 15, so it cannot establish the state described in November. It treated the acceleration threshold as an early warning, not proof that the feared outcome had materialized. Its principal published measure was Anthropic ECI, an internal adaptation of Epoch’s capability index using internal benchmark results. Those scores are not directly comparable with the public Epoch leaderboard. The report also used leading indicators whose details and trends were withheld as sensitive. At that earlier coverage date, Anthropic concluded progress remained below the threshold, with uncertainty and a measurement lag. It attributed an earlier capability jump to specific research advances that it did not believe AI had substantially accelerated. It separately found less than twofold acceleration beginning in early-to-mid 2025, crediting non-AI factors for the onset but AI tools for helping the faster pace persist through its coverage date. Distinguishing faster results from their cause is therefore part of the assessment itself.