Anthropic may have crossed its own AI acceleration warning threshold as early as July this year. In an audit published on November 6, the nonprofit evaluation group METR concluded that since October 2025, Anthropic's model test scores have been improving about 1.9 times faster, with roughly 80% of the gain attributed to AI being used in the company's own research. The audit ran for 11 weeks. METR's collaboration with Anthropic was already established — six evaluators had previously been given employee-level access — and this time the auditors also obtained benchmark results from unreleased models, training metrics, employee surveys, and internal analyses, plus interviews with about 15 researchers.
How the acceleration was measured
Behind the numbers is an abrupt change of pace. Before October 2025, Anthropic's models gained about 15 AECI composite points per year; since then, the rate has reached 29 per year — a year's progress that used to take two, a jump roughly equivalent to going from Claude 2.1 to Sonnet 4.5. The analysis covered three unreleased models that finished training between July and October 2026. METR concluded the speedup was not caused by faster growth in training compute. Employees reported that the latest internal model made them about 12 times more productive, though METR warned this is very loose evidence and may overstate the gains. AI is not yet doing fully autonomous research; the acceleration comes mainly from AI acting as an assistant, although early signs of independent contributions are emerging, and no bottleneck is visible that would slow the pace over the next 12 months.
What the redactions took away
The uncertainty around the conclusion is also significant. The true acceleration factor could be as low as 1.3 or as high as 3; counting OpenAI, Google, and xAI, the industry-wide figure falls to 1.4 — possibly because Anthropic lagged behind rivals in 2025 and a catch-up effect is skewing the numbers. Anthropic holds unilateral redaction rights over non-public information and used them heavily: most of the most informative evidence was removed, including forecasts for future models, breakdowns by training step, and the bottleneck analysis, and one chart in the report is even stamped with the words FAKE DATA. In a separate review, METR noted that Anthropic's August risk report said publicly that acceleration was under 2x, while the unredacted internal version implies 1.97; the company itself concedes the threshold may have been crossed, and METR considers July 15 the likely date. That lines up with Anthropic's recent S-1 revision acknowledging that acceleration research may hit a self-improvement threshold.
METR currently agrees that acute catastrophic risk is low, but does not rule out another sharp acceleration in the next six months that could quickly push the risk rating higher. The planned check on how compute is split between capability research and safety research was halted midway by Anthropic, which said it wanted to focus on risk assessment first. Daniel Kokotajlo, author of the AI 2027 scenario, criticized the move on X, arguing that auditing compute allocation is the foundation for making slow-down commitments verifiable. The next question is whether Anthropic restores that part of the audit, and what its responsible scaling policy will trigger now that the warning line may have been crossed.
