What happened
On November 5, 2026, METR published a review of the “Risks from automated R&D” section of Anthropic’s public August 2026 Risk Report. METR says the review was part of a high-access assessment: Anthropic shared a private version of the report and other non-public data. METR also says it had the right to disclose whether Anthropic exercised redaction rights and whether important conclusions depended on redacted information; it describes relevant redactions in the review.
Anthropic’s Responsible Scaling Policy v3.4 sets two thresholds for “Automated R&D in key domains.” One is a full-substitution threshold: a model can substitute for Anthropic’s research scientists and research engineers at competitive cost, defined as within a factor of five. The other is a two-times acceleration threshold: dramatic acceleration of AI progress likely related to automated AI research and development. METR notes that Anthropic treats the acceleration threshold as an early warning, not proof that the threat has materialized.
Anthropic argued that the full-substitution threshold had not been crossed because researchers still do work that models do not replace, despite the company’s incentive to substitute models where possible. Its report also presented benchmark evidence, which METR says it considers less important and more uncertain than the direct observation that staff remain necessary. On acceleration, Anthropic reported meaningful acceleration beginning in early-to-mid 2025 but below a factor of two. It attributed much of the 2025 acceleration to factors other than its use of AI, while also saying AI had helped faster trends continue through the report’s coverage date. It described weak evidence for a sustained acceleration in its AECI measure.
METR agrees that, as of July 15, internally deployed models posed low risk of directly automating R&D in key domains or causing acute catastrophic risks from acceleration. But it says Anthropic’s conclusion that the early-warning thresholds had not been crossed rests on contestable judgment calls that are not communicated clearly in the public report. METR’s researchers and Anthropic staff lean toward the threshold not having been crossed; METR nevertheless considers crossing plausible and says the report understates the amount of acceleration. The source does not describe this as evidence that the underlying catastrophic threat was already high.
What the archive adds
This is a new review in a developing story about how Anthropic defines and communicates R&D risk thresholds. Our earlier review of the high-stakes misalignment section questioned whether Anthropic’s methodology would reliably detect higher risk. A separate METR companion-post note covered process disclosure; this review adds a specific disagreement about automated R&D acceleration.
Trend signal
This supports the automated R&D threshold analysis: Anthropic’s RSP makes a capability threshold part of its public risk framework, but applying the threshold depends on judgments about substitution, acceleration and attribution. METR’s review shows that external evaluators can agree on low immediate risk while disputing whether the warning condition was met.
Why now
METR published this review after Anthropic’s August report and frames it as part of its broader high-access assessment of AI R&D acceleration. Its November publication brings the public disagreement into view, while distinguishing the review of the report from a separate holistic assessment.
Why it matters
Teams building or evaluating advanced AI systems should not collapse a threshold status into a single risk label. Here, “low” describes METR’s assessment of direct catastrophic risk at a specified date; the early-warning threshold is a separate monitoring mechanism. Conflating them can make a warning framework look either more alarming or more reassuring than it is.
The review also highlights how much a public assessment depends on measurement and attribution choices. “Less than 2x” acceleration is not self-explanatory unless readers can inspect the indicators, baseline, causal reasoning and uncertainty. For organizations adopting similar thresholds, the useful practice is to publish what observations would trigger a threshold and which judgments remain contestable, rather than presenting a binary conclusion without its uncertainty. METR’s disagreement does not establish that acceleration is dangerous; it does show why independent scrutiny of definitions and evidence matters.
Limits and uncertainty
This is METR’s review of one section of Anthropic’s report, not a definitive measurement of AI-driven research acceleration. The source describes access to non-public material, but readers of the public review cannot independently inspect all that material. METR says both its researchers and relevant Anthropic staff lean toward the threshold not having been crossed, while judging that crossing remains plausible. It does not establish that the threshold was crossed, that models could fully substitute for research staff, or that catastrophic risk was high on July 15. Its conclusions rely in part on judgments whose uncertainty METR says Anthropic’s public report did not communicate clearly.
Reader takeaway
Anthropic’s “low” risk conclusion and METR’s concern about a plausible early-warning-threshold crossing are not contradictory: they answer different questions. Read the threshold definition, evidence and uncertainty alongside the headline risk level.