Anthropic may have crossed its AI acceleration threshold
METR estimates that Anthropic’s model improvement pace has doubled, with about 80% of this acceleration linked to AI use in its research. The group had set a warning threshold around a doubling.
The pace of Anthropic’s model progress may have crossed its own AI acceleration warning threshold as early as July, according to an audit conducted by the nonprofit METR and published on November 6, 2026. The organization estimates that the development speed nearly doubled, with a significant portion driven by the use of AI tools in internal research. This issue matters because such a snowball effect fuels fears of ever faster model improvements, making regulation more difficult.
An external audit, but with blind spots
Between August 12 and October 29, 2026, METR had access to Anthropic’s internal data for eleven weeks, conducting about 15 interviews with the group’s researchers. The nonprofit says it worked on Anthropic devices for security reasons and received no funding from the company. According to its report, this mission gave it rare access to benchmark results, training metrics, internal surveys, and unpublished analyses.
But the exercise has its limits. Anthropic had redaction rights over non-public information and, according to METR, removed much of the most useful material, including forecasts on future models and bottleneck details. As a result, the public reading relies more than expected on METR’s judgment, which the organization admits almost verbatim by explaining this leaves the public dependent on its judgment “much more than would be ideal.” A phrase that is hardly reassuring, though at least honest.
Speed gains mostly linked to AI in research
METR concludes that Anthropic’s improvement speed increased by about 1.9 times since October 2025, based on the AECI, a composite indicator aggregating several benchmarks. Before October 2025, models gained about 15 AECI points per year; since then, the pace rose to 29 points per year. The gap is comparable, according to the audit, to the jump from Claude 2.1 to Sonnet 4.5 in a single year instead of two.
The nonprofit also estimates that about 80% of this acceleration comes from AI used as an assistant in Anthropic’s research. An employee quoted in the report summarizes the situation by saying: “Now only bottlenecked on ideas & compute, no longer bottlenecked on code.” Other testimonies mention experiments once long to set up and now almost free. Conversely, METR judges these self-assessments of productivity very fragile and probably too optimistic.
Why Anthropic’s warning threshold becomes tricky
The issue is not only statistical. In its Responsible Scaling Policy, Anthropic considers that a doubling of progress speed due to research automation should trigger a warning level. METR thinks this threshold may have been reached on July 15, 2026, or at least dangerously approached. The organization also says Anthropic acknowledged the threshold crossing as “plausible” in its unredacted internal version, while the public August 2026 version spoke of an acceleration “less than a factor of 2.”
METR nevertheless shares the view that an acute catastrophic risk remains “low” at this stage. The problem is that the audit does not exclude a new runaway within the next six months. In short, the machine is not yet autonomous, but it is already gaining speed with the help of the machine. This kind of feedback loop is what security teams want to avoid, precisely because it can disrupt control timelines.
Disputed figures, divergent voices within Anthropic
METR highlights a significant margin of uncertainty: the real acceleration factor could be as low as 1.3 or as high as 3. Including OpenAI, Google, and xAI in the comparison would reduce the aggregated figure to 1.4, possibly because Anthropic mainly caught up on a 2025 lag. Within the company itself, opinions also diverge: one researcher attributes only 15 to 20% of productivity gains to AI, while another sees new hires as the most important factor.
Anthropic was also to allow METR to verify the distribution of its compute between capability research and safety research, but this part was suspended midway. The company said it wanted to focus first on risk assessment before returning to compute control. Daniel Kokotajlo, author of the “AI 2027” scenario, criticized this choice on X, arguing that compute auditing is the basis for making slowdown commitments truly verifiable.
Key points
- METR audited Anthropic for eleven weeks, from August to October 2026.
- The progress pace nearly doubled since October 2025.
- About 80% of the acceleration came from AI in research.
- Anthropic considers the threshold crossing plausible.
- The threshold may have been reached on July 15, 2026 [to be verified].
- The comparison also includes OpenAI, Google, and xAI.
By the numbers
- 11 weeks — duration of METR’s audit at Anthropic, August–October 2026.
- 1.9 times — estimated acceleration of progress pace since October 2025.
- 80% — estimated share of acceleration linked to AI in internal research.
- 15 to 29 AECI points per year — pace evolution before and after October 2025.
- 1.4 — aggregated acceleration when including OpenAI, Google, and xAI.