METR finds ~2× AI R&D acceleration inside Anthropic

METR published an independent assessment on 5 November 2026 of how much AI is accelerating AI research and development inside Anthropic. Staff worked on-site with high access from 12 August to 29 October 2026. METR says this engagement is distinct from the 15 October independent-evaluators announcement covered in our employee-level access post.

What METR reports

On Anthropic’s AECI benchmark index (Anthropic’s version of the Epoch Capabilities Index), METR estimates scores have been growing at 1.9× the pre-October 2025 rate since October 2025, with a 95% confidence interval of 1.5-2.8×. In plain language, METR says Anthropic appears to have made roughly two years of historical benchmark progress over the last year.

Separately, METR’s subjective staff view is that as of October 2026 Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate it did between March 2024 and October 2025, with high uncertainty; METR would not be surprised if the true figure were closer to 1.3× or 3×.

METR states it is confident the acceleration was not due to an acceleration in the scaling of training compute. Interviews and surveys are described as consistent with AI assistance (rather than fully autonomous contributions) playing a large role, with some signs that more autonomous contributions could matter soon.

Redactions

Anthropic had rights to redact non-public information. METR says it could freely share its highest-level risk conclusions, but Anthropic redacted a majority of the evidence METR found most informative for those conclusions. Anthropic described the redactions as aimed at removing confidential information that could accelerate other AI developers. METR notes that much of the redacted evidence has multiple plausible readings, so the public cannot fully judge the evidence for itself.

Anthropic’s reply on X

Anthropic co-founder Sam McCandlish (@Sam_McCandlish) replied on X defending the redactions and arguing that models still cannot autonomously handle most critical training work, while acknowledging that AI speeds research. He also pointed to an Anthropic progress / RSP write-up coming in the following weeks. We summarise that reply from triage of the public post; wording here is paraphrase, not a direct quote.

Why this matters

If you build or buy frontier models, this is one of the few independent, high-access reads of whether AI is already bending a lab’s own R&D curve, and METR’s answer is a cautious yes on well-specified tasks without a training-compute story. The redaction wall means outsiders still cannot audit the most informative evidence, so treat the multiples as METR’s judgment under constraint, not as a public dataset you can re-fit. Pair it with the earlier access-tier story and with other evaluator work such as METR and Redwood’s OpenAI agent-monitoring red team rather than leaping to recursive self-improvement claims this report does not make.