METR publishes independent assessment estimating Anthropic's AI capability progress is about 2x the historical rate, largely due to AI-assisted R&D

METR, an independent AI evaluator, published a report from a high-access evaluation inside Anthropic, run from 12 August to 29 October 2026. It estimates Anthropic's model capabilities on well-specified tasks are improving at roughly 2x the March 2024 to October 2025 rate, with a wide uncertainty range of about 1.3x to 3x. Growth on AECI, Anthropic's version of the Epoch Capabilities Index, since October 2025 is put at 1.9x (95% CI 1.5 to 2.8x). METR found no evidence the speed-up came from faster compute scaling and saw no significant bottlenecks over the next 12 months, though it says it is not confident in that. Anthropic redacted some key findings.

This is rare: an outside party with real access to a frontier lab measuring how fast it is progressing, rather than the lab reporting on itself. It gives regulators, safety researchers and investors a number to argue over, and it supports the view that AI-assisted R&D is already speeding up development. It does not change anything deployed today. The practical effect is on policy and safety debates, such as how much warning time there is before risky capabilities appear. The wide range and the redactions limit how far anyone can rely on the headline 2x.

"2x acceleration" applies to well-specified tasks on one lab's own benchmark index, so it should not be read as general progress or as proof of a runaway feedback loop. The claim that further progress could bring even faster acceleration is a forecast resting on an absence of observed bottlenecks, which METR itself holds with low confidence. Redacted findings and the lack of linked primary documents in our extraction mean outsiders cannot yet check the key claims, and the source is a forum post (LessWrong) relaying the report.

1. 01Publication of the full METR report and any unredacted detail, and whether other labs allow comparable high-access evaluations.
2. 02Independent measures of the same trend, such as Epoch AI's Capabilities Index and METR's task-length results for new Anthropic models, to see if growth stays near 1.9x or falls within the 1.3x to 3x range.
3. 03Anthropic's response and any policy or safety framework changes in the next quarter, such as revised capability thresholds or regulator citations of the report.