METR: At Anthropic, the Pace of AI Progress Has Roughly Doubled, Mostly Driven by AI Itself

METR, a US-based independent evaluation organization for AI systems, released on November 5, 2026, the public summary of an audit conducted inside Anthropic: is AI already speeding up AI research and development at the Claude developer? The answer, delivered with significant caveats and following extensive redactions, is yes—by roughly a factor of 2 on well-defined tasks, largely driven by the internal use of models.

The announcement complements the passage in Anthropic’s amended S-1/A prospectus in which the company noted it could “imminently” approach the acceleration threshold defined in its Responsible Scaling Policy.

What METR Claims to Have Seen

From mid-August to late October, a METR team (roughly 2.5 full-time equivalents) had access to internal data: internal benchmark and AECI scores, pre-training and RL metrics, training compute by domain, uplift surveys, interviews with around fifteen staff members, and velocity analyses. Anthropic retained the right to decline requests and subsequently redacted a large share of the public report. METR received no payment; Anthropic provided computers and tokens.

Three subjective conclusions frame the document:

1. Pace — As of October 29, 2026, Anthropic is likely improving model capabilities on well-defined tasks at roughly double the pace observed between March 2024 and October 2025. METR would not be surprised if the true figure fell around 1.3× or 3×.
2. Attribution — A majority of the current acceleration is likely due to AI used for R&D (subjective order of magnitude: ≈ 80%), even if other factors contributed in late 2025.
3. Bottlenecks — METR did not observe a clear bottleneck (algorithms, training compute, experiment compute, data) likely to stall further acceleration over the next 12 months, though with limited confidence.

On the AECI index, scores accelerated by an estimated factor of 1.9× since October 2025 (95% CI: 1.5–2.8×). In other words, one recent year represents about two years of historical progress on this index. METR emphasizes that the acceleration does not stem from an acceleration in training compute—a point it expresses confidence in, even though the supporting details were redacted.

Internal interviews and surveys point in the same direction, but METR treats them with caution: self-reported Slack polls show an average uplift of up to roughly 12× in October 2026 for the most capable internal model, evidence deemed “very weak.” Fully autonomous AI contributions to the most critical parts of the training pipeline are not yet substantial, though “signs of life” are emerging.

What Anthropic Redacted

The public report acknowledges it right from the introduction: Anthropic removed a majority of the evidence that METR considered most informative, citing the protection of confidential details that could aid competing developers. METR retained the right to disclose this. The accepted outcome: the public must rely far more on METR’s judgment than on transparent underlying evidence.

Among the redacted material are detailed estimates of acceleration and attribution by pipeline stage, portions of the bottleneck analysis, and projected capability metrics for future models. Unpublished internal models (referred to as “Model 2,” “Model 3,” and “Model 4” in the text) are included in the AECI trends: this is partly why the estimated acceleration (1.9×) exceeds that from the Haiku 5.5 system card (1.7×).

Anthropic also canceled midway through a planned evaluation track focused on verifying compute allocation (capabilities vs. safety), prior to any substantive findings, in order to refocus the engagement on risk assessment.

The Review of the August Risk Report

On the same day, METR published a review of the “Risks from automated R&D” section of Anthropic’s August 2026 Risk Report, having had access to the non-public version. The two documents should be read in tandem.

METR agrees with Anthropic that, as of July 15, 2026, the catastrophic risk stemming from direct R&D automation was low. Conversely, it considers it plausible that the RSP’s early warning threshold of 2× had already been crossed by that date, and argues that the public report understated both the level of acceleration and the uncertainty conveyed privately (a private metric indicating roughly 1.97× was summarized publicly as “less than a factor of 2”).

What Remains Open

• Redacted evidence: Outside observers cannot independently reconstruct the detailed reasoning.
• AECI ≠ real-world impact: METR notes that the index primarily measures well-defined, saturating tasks; real-world impact outside the lab could differ significantly.
• What’s next: METR sees a significant risk that acceleration will intensify further, though hyperbolic growth cannot be proven solely from public AECI data.