What happened
In an October 14, 2026 research note, METR described an assessment of four questions: how quickly internal model capabilities are improving and whether that rate has accelerated; how much any acceleration is attributable to AI automation rather than other factors; what bottlenecks could limit near-term growth; and whether the proposed methods would reliably reveal a more dramatic acceleration if one occurred.
For capability progress, METR says an assessor should not rely on a single benchmark. The proposed evidence includes scores across capability benchmarks and aggregate measures such as the Effective Compute Index, intermediate measures such as pretraining loss, and researchers’ perceptions of how internal models are changing. The note calls for benchmark details, scoring methods and sample model answers, as well as historical records of progress in pretraining, post-training and other parts of development. It also suggests checking how developers believe progress in those domains contributes to overall capabilities.
To investigate the role of AI, METR proposes comparing multiple lines of evidence because the “no AI” counterfactual is difficult to establish. These include AI-agent inference spending by team and over time; researchers’ accounts of how models contributed to research; and productivity measures, including crude indicators such as lines of code per researcher-day or self-reported uplift. The assessor would also need dated estimates of algorithmic progress, and inputs that could independently explain faster results: team labor, training and experimental compute, data, and other important resources. METR says those inputs should be broken down by team or function where relevant.
The plan also calls for evidence about constraints and forecasts: information on possible bottlenecks such as training compute or human data, existing projections of top-line metrics, and interviews with researchers in pretraining, reinforcement learning and infrastructure. METR says access should cover the longest possible timeframe, with precise dates, and that data could be analyzed on developer hardware or onsite, requiring only minimal access to developer infrastructure. The note does not announce a completed assessment or a participating developer.
What the archive adds
New Runtime’s earlier coverage on October 23, 2026 reviewed the high-stakes misalignment section of Anthropic’s August Risk Report, including questions about whether its methodology could reliably detect higher risk. The October 14 METR note is a new, related methodological proposal: it specifies what evidence an assessor would need to investigate AI-driven R&D acceleration, rather than reporting an audit result.
Trend signal
This supports the concern behind New Runtime’s verification-ownership shift: consequential claims need evidence that can be checked, not just an assessor’s conclusion. METR applies that idea to a difficult causal question by proposing multiple measures, historical data and direct researcher interviews. The note is a framework, not proof that those checks have been performed.
Why now
METR frames the assessment around whether capabilities progress has accelerated recently and whether AI automation contributed. Its proposal emphasizes dated observations and a historical baseline, making the timing relevant to distinguishing an actual change in trajectory from growth in other inputs such as compute or labor.
Why it matters
For teams building with AI, the proposal is a reminder that productivity claims and capability forecasts are hard to interpret without a counterfactual and a record of changing inputs. A faster research cycle might reflect more compute, larger teams, better data, algorithmic advances, AI assistance, or some combination. Reporting only model scores or anecdotes can obscure those differences.
METR’s checklist also makes the operational burden visible. A credible review would need internal-only model results, dated resource data and access to researchers across several functions. For organizations that want external scrutiny, the practical question is not simply whether an evaluator is independent, but whether the evaluator can inspect enough of the underlying evidence to test competing explanations. The proposal offers a basis for discussing that access before a high-stakes conclusion is published.
Limits and uncertainty
The note sets out what an assessor would need; it does not report measurements, quantify any acceleration or attribute a change to AI. METR does not identify a developer participating in such an assessment or say that the proposed data access has been secured. Several measures are inherently difficult to interpret: benchmark performance may not track real-world usefulness, productivity estimates depend on uncertain counterfactuals, and forward-looking forecasts are highly uncertain. The proposal therefore cannot establish whether AI is currently speeding up AI R&D or how reliably a future assessment would detect a dramatic change.
Skill to build now
If your team makes claims about AI improving its own R&D, start a dated input-and-output log: model capability measures, AI-agent usage or inference spend, relevant compute and staffing, and specific examples of AI contributions. Keep the measures disaggregated by team or stage where possible. That will make later reviews more useful than a retrospective estimate based on memory alone.
Reader takeaway
METR has published a blueprint for investigating AI-driven acceleration of AI research, not evidence that such acceleration has happened. Its key test is whether capability trends can be connected to AI’s contribution while accounting for compute, labor, data and other causes.