Anthropic: METR Signals a Threshold Alert Crossed in July

METR claims that internal documents from Anthropic imply an acceleration close to 1.97 and considers it possible that the warning level was reached on July 15. The organization laments significant redactions and Anthropic's halt of an audit on compute allocation. Despite this, METR currently shares the assessment of a "low" acute catastrophic risk, while not ruling it out if the acceleration continues in the coming months.

Anthropic exercises its right to redact most information

Anthropic had a unilateral right to redact non-public information, and these rights have been extensively used, according to METR. The organization asserts that the majority of the most informative elements have been removed, including model forecasts, training breakdowns, and bottleneck analyses. Anthropic presents these removals as a measure to protect confidential information that could accelerate other AI developers. METR writes that the public must rely on its judgment far more than desirable and that a transparent discussion of the evidence is impossible; a graph in its report displays "FAKE DATA" in large letters. Meanwhile, Anthropic has canceled an audit on compute allocation between capabilities and safety, explaining that it wants to focus first on risk assessment and return to it later. In a supplementary review, METR also examined the risk report from August 2026: the public version indicates an acceleration "below 2," while the internal version would imply 1.97, and the company considers it "plausible" that the threshold has been crossed. METR accuses the public report of downplaying the acceleration and its uncertainty and believes it is possible that the warning level was reached on July 15.

Measured acceleration: 1.9x since October 2025 on the AECI

Since October 2025, METR has observed a progression of test scores approximately 1.9 times faster than between March 2024 and October 2025, relying on the AECI, a composite indicator. According to the organization, Anthropic has condensed in one year advancements that previously took two years, in a magnitude comparable to the leap from Claude 2.1 to Sonnet 4.5. Three unpublished models, whose training phase ended between July and October 2026, are included in this analysis. Before October 2025, the annual increase in AECI points for the models was about 15; since that date, it would have risen to 29.

Role of assistant AI and self-reported productivity

METR believes that the acceleration is not due to a faster increase in training compute. Auditors attribute about 80% of the gain to the use of AI in research, with AI primarily acting as an assistant. An employee indicates that work today is constrained by thought and computing power rather than programming, while others report that experiments that previously required one to two days of setup are now "almost free." In informal surveys conducted on Slack in October 2026, staff members indicated they were on average about twelve times more efficient with the latest internal model. However, METR considers these self-assessments to be very rough, not representative, and likely exaggerated; the estimates mentioned vary from 1.2x with Claude 3.7 Sonnet to 12x with an internal model 4. According to METR, fully autonomous research is not yet a reality, although there are hints that more independent contributions could soon carry more weight.

Contested estimates and a weaker signal from the sector

METR highlights significant uncertainty: the actual acceleration factor could be between 1.3 and 3. When including OpenAI, Google, and xAI, the estimate drops to 1.4 for the sector, which METR links to a catch-up effect from Anthropic after a delay in 2025. Internally, opinions differ: one researcher reduces the share of AI in productivity gains to 15 to 20% and is wary of self-reports, while another emphasizes the impact of new hires in recent months.

Access framework and scope of collaboration with METR

From August 12 to October 29, 2026, a METR team of 2.5 full-time equivalents had unprecedented access to unpublished internal benchmarks, training metrics, internal surveys, and analyses, along with about 15 interviews with researchers. METR specifies that it received no funding from Anthropic and worked on the company's devices for security reasons. Dario Amodei had committed to involving external auditors, and in October, Anthropic announced a collaboration granting METR access "similar to that of employees." However, METR presents the ongoing audit as a separate project, with a narrower scope, continuing its mission to assess risks of loss of control over systems.

Current risk deemed "low," continued acceleration possible

METR has identified no bottlenecks likely to slow the pace over twelve months and believes there is a significant risk of continued acceleration. The organization currently shares Anthropic's assessment of a "low" acute catastrophic risk, while not ruling out a rapid acceleration in the next six months that could raise the risk score. METR was also to audit the allocation of compute between capabilities and safety; Daniel Kokotajlo argues that this exercise constitutes the basis for verifiable commitments to slow down. In its internal policy, Anthropic stipulates that a doubling of the pace of progress triggers a warning level.