METR measured Anthropic’s AI R&D pace at 1.9x — and Anthropic redacted most of the evidence behind the number

On 2026-11-05, METR (Model Evaluation and Threat Research) published the result of a high-access evaluation it conducted inside Anthropic from August 12 to October 29, 2026. The findings are a benchmark for the industry: Anthropic’s capabilities, measured on AECI (Anthropic’s version of the Epoch Capabilities Index), have been growing since October 2025 at an estimated 1.9 times the pre-October-2025 rate, with a 95% confidence interval of 1.5x to 2.8x. METR characterizes this as “roughly two years of historical progress over the last year.”

That is the number. The story, however, is what happened to it.

Anthropic held contractual rights to redact non-public information from the report and exercised them to redact a majority of the evidence METR found most informative, describing the redactions as “aimed at removing confidential information that could accelerate other AI developers.” METR’s own redaction summary states the consequence plainly: the report “does not achieve transparency about evidence,” meaning “the public has to rely on METR’s judgment much more than would be ideal, given the weight of the questions answered here.”

We are left with a structural irony: the most consequential external measurement of AI safety in the current cycle was produced by an independent party, but the primary evidence behind it was withheld by the subject of the measurement.

The Governance Gap

This report surfaces the core tension in the “embedded evaluator” model that frontier labs are now rushing to adopt. On the same day the METR report was released, Anthropic announced a major partnership with Accenture—led by its specialist AI business, Faculty—to embed evaluators with employee-comparable access inside the company. Anthropic and Accenture each expect to invest at least $1 billion over the next five years, with Anthropic funding Accenture’s work directly.

While Anthropic describes this as a move to make accountability “more verifiable,” the company also acknowledges the systemic challenges. Anthropic explicitly states that there are “as yet, no standards for what information embedded evaluators should have access to, or how they should report what they find,” and no settled funding system for truly independent evaluation. Anthropic maintains that long-term, this should be pooled or government-sourced, as outlined in its Advanced AI Framework from June.

The result is a fragile ecosystem. METR notes that both it and Anthropic researchers have significant uncertainty about how to synthesize evidence into a holistic picture of acceleration, and the independent scrutiny of the process—including critiques regarding METR’s own funding ties to Anthropic and Accenture’s commercial partnership—suggests that under current frameworks, the conditions for adversarial auditing are hard to satisfy.

The Bottleneck is Auditability

METR’s report is not an alarm on specific dangerous capabilities, but a diagnostic of R&D velocity. METR concluded it is confident the observed acceleration was not due to an increase in the scaling of training compute, and it did not observe significant bottlenecks in algorithmic, training, or data progress over the coming 12 months.

The fundamental takeaway for the industry is clear: the ability to measure progress is becoming as critical as the progress itself. “How quickly AI progress is accelerating has many important safety implications, since it determines how quickly the risks from advanced models increase,” METR observes.

For now, the public must contend with a report that is more of a signal than a record. Anthropic has promised its own report in the coming weeks, but the auditability question the site has been circling is now concrete: it is no longer just “should models be audited,” but “who pays the auditor, and does the evidence survive the audit.” Until these questions are settled, we are watching the most important R&D race in the world through a screen of redactions.