METR finds Anthropic may have hit its own AI acceleration warning threshold

The audit of Anthropic's progress

The nonprofit organization METR recently completed an eleven-week audit of Anthropic's internal research data. This investigation aimed to verify whether the company's AI development is accelerating at a rate that could pose significant safety risks. METR received access to unreleased model benchmarks, training metrics, and employee surveys to conduct this review.

The findings suggest that the rate at which Anthropic's models improve has nearly doubled since October 2025 compared to the previous year. This measurement is based on the AECI, which is a composite metric that bundles many different benchmarks together. Essentially, the company achieved in one year what previously took two years of development.

METR notes that this jump is roughly comparable to the progress seen between Claude 2.1 and Sonnet 4.5. The analysis looked at three models that wrapped up training between July and October 2026. The audit suggests that the pace of improvement has hit 29 AECI points per year, up from 15 points in the preceding period.

AI-driven research feedback loops

A primary concern in the AI safety debate is the possibility of a feedback loop where models improve their successors faster and faster. METR estimates that the vast majority of this current speedup is driven by AI being utilized in Anthropic's own research. This means the models are actively assisting in the creation of better versions of themselves.

One Anthropic employee noted in an interview that the team is no longer bottlenecked on code. Instead, the primary constraints are now ideas and compute. Other employees reported that experiments that previously required one to two days of preparation are now essentially free to run because of AI assistance.

While AI is not yet performing fully autonomous research, there are signs that independent contributions from models may soon play a larger role. METR found no current bottlenecks that would slow this pace over the next twelve months. This suggests a significant risk that the acceleration will continue to grow over the coming year.

Crossing the safety threshold

Anthropic has a Responsible Scaling Policy that identifies specific risk indicators for AI development. One of these indicators is a doubling of the pace of AI progress caused by automated AI research. METR's findings suggest that Anthropic may have crossed this warning threshold as early as July 15, 2026.

In its own public risk report from August 2026, Anthropic stated that the acceleration was less than a factor of 2. However, METR found that the internal, unredacted version of that data implies a value of 1.97. METR accuses the public report of understating both the actual acceleration and the level of uncertainty surrounding the data.

The organization currently shares Anthropic's view that the acute catastrophic risk remains low. However, they warn that another sharp acceleration over the next six months could quickly push that risk rating higher. The audit highlights how quickly the situation could change if the feedback loop continues to intensify.

Data redactions and transparency

Anthropic maintained unilateral redaction rights over all non-public information provided to METR. The nonprofit reported that the majority of the evidence they rated as most informative was stripped out of the final report. This included forecasts for future models, breakdowns by training steps, and specific bottleneck analyses.

Anthropic stated that these redactions were intended to remove confidential information that could help other AI developers. Because of this, METR notes that the public must rely on their judgment much more than would be ideal. One graph in the METR report even features the words "FAKE DATA" in large letters to highlight these gaps.

The lack of transparency makes it difficult for outside observers to verify the exact scale of the acceleration. While METR provides a specific estimate, the actual growth factor could be as low as 1.3 or as high as 3. This range demonstrates how much of the internal data remains hidden from public view.

Industry context and next steps

When looking at the broader industry, the acceleration number drops to just 1.4. This lower figure may be because Anthropic lagged behind competitors like OpenAI, Google, and xAI in 2025. The current data might be skewed by a catch-up effect as Anthropic tries to close the gap with its rivals.

Internal views at Anthropic are not uniform regarding these productivity gains. One researcher believes AI's share of the gains is only 15 to 20 percent and distrusts the self-reports of colleagues. Another researcher argued that new hires are actually the single most important factor in the recent months of progress.

METR was originally intended to audit how Anthropic splits its compute between capability research and safety research. However, Anthropic canceled that portion of the project to focus on risk assessment first. Critics argue that auditing compute allocation is the foundation for making verifiable commitments to slow down development.