METR published its Anthropic assessment on 5 November 2026, following evaluation from 12 August to 29 October. Final public report submitted: 3 November. Assessment process.
Read the rate, then the uncertainty
Since October 2025, Anthropic’s AECI benchmark-progress rate was an estimated 1.9× its earlier pace (95% confidence interval: 1.5–2.8×). This measures progress, not doubled capabilities or proven AI causation. METR separately judges AI likely explains most current acceleration, with substantial uncertainty. Findings and interpretation.
Keep three questions separate when reading a headline like this: what changed on the measurement, what explains the change, and whether it helps the work you need done. An estimate about a developer’s research trajectory does not answer a team’s deployment question by itself.
Check what is available to inspect
METR says much key evidence was redacted, limiting public verification; Anthropic cited confidentiality. METR received devices and tokens, but no payment. Evidence and disclosures.
Record that boundary in any internal summary. Do not fill an unavailable evidence gap with either an assurance or an accusation. Keep the reported result, the evaluator’s judgment and your own assumptions visibly separate.
Use it to sharpen a local trial
For a team considering AI-assisted research or engineering, a useful next step is a small comparison with written acceptance criteria:
- Choose an outcome. Define a completed, reviewed result rather than counting generated suggestions.
- Keep the comparison legible. Record task mix, tools, model version and who reviewed the work.
- Count the whole job. Include rejected attempts, correction time, review effort and costs.
- Write the decision rule first. Decide which quality checks must pass and what would justify wider use.
These are proposed evaluation steps, not results of a trial we ran. Use local evidence to decide whether a workflow improves, rather than importing a headline multiplier into a staffing or delivery forecast.