Anthropic may have reached its own warning threshold for AI-driven research acceleration as early as July 15, 2026, an external review by nonprofit METR has found. The assessment estimates that the pace of improvement in Anthropic’s models has nearly doubled, with AI assistance accounting for most of the increase, but substantial uncertainty and extensive redactions limit the public evidence.
The finding concerns a risk indicator in Anthropic’s Responsible Scaling Policy: a doubling in the pace of AI progress caused by automated AI research. Such acceleration matters because increasingly capable systems could help develop their successors, creating a feedback loop that makes future capability gains harder to anticipate or manage. METR’s findings do not establish that this kind of autonomous cycle is already operating.
METR examined internal research information from August 12 to October 29, with a team equivalent to 2.5 full-time staff. Its access included benchmark results for unreleased models, training metrics, employee surveys and internal analyses, supplemented by roughly 15 interviews with Anthropic researchers. METR received no payment from Anthropic and used company devices for security reasons.
The central measurement used AECI, a composite score combining multiple benchmarks. Anthropic’s models improved at approximately 29 points per year after October 2025, compared with about 15 points annually between March 2024 and October 2025. That represents an estimated 1.9-fold increase in the rate of progress. The analysis included three unreleased models that completed training between July and October 2026.
METR concluded that faster growth in training compute did not explain the acceleration and estimated that AI use within Anthropic’s research accounted for about 80 percent of the speedup. The gains came primarily from systems assisting researchers rather than independently conducting research. Interviews described coding and experiment preparation becoming less restrictive, leaving ideas and compute as more important constraints.
Employee productivity estimates provided additional, weaker evidence. In informal internal polls during October, employees reported average productivity gains of roughly twelvefold when using the latest internal model. METR cautioned that these responses were unrepresentative and likely overstated actual improvements. Researchers also disagreed about the causes: one attributed only 15 to 20 percent of productivity gains to AI, while another viewed new hires as the most important recent contributor.
The overall acceleration estimate remains imprecise. METR put the possible range at 1.3 to 3 times the earlier pace. Including models from OpenAI, Google and xAI reduced the industry-wide estimate to about 1.4 times. One possible explanation is that Anthropic’s results partly reflect catching up after trailing competitors in 2025, rather than a comparable acceleration across the field.
A companion review examined Anthropic’s August risk report, which publicly described acceleration as below a factor of two. The internal version implied a figure of 1.97, and Anthropic considered it plausible that the threshold had been crossed. METR judged that the public account understated both the acceleration and the uncertainty, and considered July 15 a possible crossing date.
Public scrutiny of those conclusions is constrained by Anthropic’s control over disclosure. The company had unilateral rights to redact non-public information and removed most of the evidence METR considered most informative, including future-model forecasts, training-step breakdowns and bottleneck analysis. Anthropic said the removals protected confidential information that could help competing developers accelerate their work.
METR currently agrees with Anthropic’s assessment that acute catastrophic risk is low, but could not exclude another sharp acceleration within six months. A separate planned examination of how Anthropic divides compute between capability and safety research was canceled midway through. Anthropic said it wanted to prioritize risk assessment and return to the compute review later.
Sources: The Decoder