METR Says Anthropic AI R&D Pace Is About Twice Historical

METR published two linked assessments on 5 November 2026: an independent review of AI R&D acceleration at Anthropic as of October 2026 and a review of the automated-R&D section of Anthropic’s August 2026 Risk Report. Every acceleration figure and judgment below is METR’s, not Anthropic’s PR. Anthropic exercised heavy redaction rights over evidence METR says most informed its conclusions.

METR’s headline subjective judgment: as of October 2026, Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate seen between March 2024 and October 2025. METR stresses high uncertainty and says it would not be surprised if the true acceleration were closer to 1.3x or 3x that historical rate.

What METR thinks it can say in public

From 12 August to 29 October 2026, METR staff ran a high-access evaluation inside Anthropic. On public numbers, METR reports that since October 2025 Anthropic’s AECI (Anthropic’s version of the Epoch Capabilities Index) has grown at an estimated 1.9x the pre-October 2025 rate, with a 95% confidence interval of 1.5–2.8x — implying roughly two years of historical progress in about a year, on that index. METR says it is confident the acceleration was not due to faster scaling of training compute.

On causes, METR guesses that a majority — “perhaps ≈80%” — of the current (October 2026) acceleration is due to AI use in R&D, with substantial uncertainty and with non-AI factors more important in late 2025. METR did not identify clear bottlenecks or steep diminishing returns that would block further compounding over the next year, again with significant uncertainty. Detailed estimates of AI’s cumulative and instantaneous effects on AECI and training-pipeline stages were redacted.

Redactions and the early-warning argument

METR says Anthropic redacted a majority of the evidence METR found most informative, describing the cuts as aimed at removing confidential information that could accelerate other developers. METR discloses that redactions happened and that important conclusions rested partly on material the public cannot see — which limits how far outside readers can re-derive the ~2x view.

On the Risk Report review, METR agrees with Anthropic that by 15 July internally deployed models posed low risk of directly automating R&D in key domains or causing acute catastrophic risks from acceleration. But METR argues that the public claim of not having crossed the RSP’s “2x acceleration” early-warning threshold rests on contestable judgment calls that are not clearly communicated. While METR’s project staff and Anthropic’s relevant staff lean toward the threshold not having been crossed, METR thinks it is plausible that it had been, and that the public report understates the acceleration then under way.

Why independent assessors matter for UK readers

TSN previously covered Anthropic embedding METR staff with employee-like access and METR’s broader loss-of-control assessment offer to frontier labs. Adjacent lab-governance reporting such as OpenAI’s training-containment work shows the same theme: voluntary RSP-style thresholds only discipline markets if outside evaluators can publish enough to let governments and buyers judge them.

For UK policy and RegTech readers, the story is oversight quality. Heavy redactions mean the public AECI multiple and the ~80% AI-attribution guess are METR’s synthesised judgments under constraint, not a full evidence dump Parliament or the AI Security Institute can re-audit from open text alone. That is still more transparency than no third-party write-up — and it is why hub tagging for AI regulation fits.

What to watch after the dual METR posts

Watch next for any Anthropic response that restores redacted evidence categories, updates to RSP v3.4 threshold language, and whether other labs invite comparable public METR write-ups. Until then, treat ~2x (range 1.3–3x), AECI ~1.9x [1.5–2.8], the ~80% AI-share guess and the “threshold plausibly crossed” line as METR’s assessed judgments under heavy redaction — not Anthropic marketing, and not settled measurement science.

Uncertainty is the honest part of the brief

METR’s willingness to publish a 1.3x–3x surprise range alongside a central ~2x view is more useful to policymakers than a false precision number. The same honesty about the ~80% AI-attribution guess being a guess, and about redactions blocking public re-derivation, is what makes the dual posts usable for UK AI-security readers even when the underlying lab evidence stays sealed.

Sources

• Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 (METR, 5 November 2026)
• Review of the R&D section of Anthropic’s August 2026 Risk Report (METR, 5 November 2026)

Related stories

• Anthropic Says It Has Embedded Six METR Staff With Employee-Like Access
• METR Offers Frontier AI Labs Independent, Public Assessments of Loss-of-Control Risk
• OpenAI Training Containment (TSN)