METR Questions Anthropic’s AI R&D Risk Assessment

METR reviewed the “Risks from automated R&D” section of Anthropic’s public August 2026 Risk Report as part of an assessment that also used a private report and other non-public data shared by Anthropic. METR agrees that models deployed internally by July 15 posed low risk of directly automating research and development in key domains or causing acute catastrophic risks from accelerated progress. It says, however, that the public report understates the acceleration underway and that it is plausible Anthropic had crossed its early warning threshold by then. METR notes that its assessment relied in part on judgments that the public report did not clearly communicate.

Anthropic’s RSP v3.4 sets two conditions for its “Automated R&D in key domains” threshold: models fully substituting for all research scientists and engineers at competitive costs, defined as within a factor of 5, or a 2x acceleration in AI progress likely related to automated AI research. Anthropic argued neither condition had been met by July 15. It cited the continued work of human researchers as evidence against full substitution and said leading indicators showed acceleration beginning in early-to-mid 2025, but by less than a factor of 2. It also attributed much of the 2025 acceleration to factors other than its AI models, while saying those models helped faster trends continue through the report’s coverage date.

Anthropic characterized the threshold as an early warning, not proof that catastrophic harm had materialized, and rated the July 15 risk “low.” Its threat model describes possible future risks if AI systems automate research in fields including robotics, energy, cyberwarfare and AI itself, such as sudden shifts in national power or loss of human control. METR says the public report could have distinguished more clearly between the risk of crossing the RSP threshold and the risk of the catastrophic outcomes in that threat model.

METR finds Anthropic’s case against full substitution broadly compelling, but says confidence is lower because high-confidence benchmarks that still measure relevant capabilities are lacking. It says alignment, tools, oversight or elicitation may be limiting automation even if models have, or are close to, the needed capabilities. METR considers the public report’s conclusion on the 2x threshold overstated and says the threshold may plausibly have been crossed by July 15, although relevant researchers at both organizations generally lean toward it not having been crossed.

For leading indicators, METR calculates that figures in the private report imply 1.97 times the historical progress rate, despite the public report’s wording “less than a factor of 2.” METR says its own estimates range from 1.3 to 2.8, depending on definitions and time period, with substantial methodological uncertainty. It says “around 2” would better describe the private report’s estimate.

On attribution to AI, METR says the public report suggests greater confidence that non-AI factors explain much of the acceleration than the private report supports. The private report called it plausible that AI-attributable acceleration exceeded a factor of 2 overall by July 15; METR broadly agrees with that assessment while emphasizing that estimating the counterfactual is difficult. METR’s preferred wording says AI models were an increasingly important factor in continuing faster trends, and that overall progress was probably slightly less than twice as fast as without AI assistance, though exceeding 2x remained plausible.

METR also challenges the public report’s treatment of AECI evidence. It says Anthropic omitted private-report qualifications that later snapshots of Model 2 were expected to improve more than the public report suggested; subsequent snapshots improved AECI by about as much as Claude Mythos 5 improved on Mythos Preview. The provided review text ends while METR is disputing Anthropic’s emphasis on weak evidence for a one-time AECI jump.

METR reviewed the “Risks from automated R&D” section of Anthropic’s public August 2026 Risk Report as part of an assessment that also used a private report and other non-public data shared by Anthropic. METR agrees that models deployed internally by July 15 posed low risk of directly automating research and development in key domains or causing acute catastrophic risks from accelerated progress. It says, however, that the public report understates the acceleration underway and that it is plausible Anthropic had crossed its early warning threshold by then. METR notes that its assessment relied in part on judgments that the public report did not clearly communicate.

Anthropic’s RSP v3.4 sets two conditions for its “Automated R&D in key domains” threshold: models fully substituting for all research scientists and engineers at competitive costs, defined as within a factor of 5, or a 2x acceleration in AI progress likely related to automated AI research. Anthropic argued neither condition had been met by July 15. It cited the continued work of human researchers as evidence against full substitution and said leading indicators showed acceleration beginning in early-to-mid 2025, but by less than a factor of 2. It also attributed much of the 2025 acceleration to factors other than its AI models, while saying those models helped faster trends continue through the report’s coverage date.

Anthropic characterized the threshold as an early warning, not proof that catastrophic harm had materialized, and rated the July 15 risk “low.” Its threat model describes possible future risks if AI systems automate research in fields including robotics, energy, cyberwarfare and AI itself, such as sudden shifts in national power or loss of human control. METR says the public report could have distinguished more clearly between the risk of crossing the RSP threshold and the risk of the catastrophic outcomes in that threat model.

METR finds Anthropic’s case against full substitution broadly compelling, but says confidence is lower because high-confidence benchmarks that still measure relevant capabilities are lacking. It says alignment, tools, oversight or elicitation may be limiting automation even if models have, or are close to, the needed capabilities. METR considers the public report’s conclusion on the 2x threshold overstated and says the threshold may plausibly have been crossed by July 15, although relevant researchers at both organizations generally lean toward it not having been crossed.

For leading indicators, METR calculates that figures in the private report imply 1.97 times the historical progress rate, despite the public report’s wording “less than a factor of 2.” METR says its own estimates range from 1.3 to 2.8, depending on definitions and time period, with substantial methodological uncertainty. It says “around 2” would better describe the private report’s estimate.

On attribution to AI, METR says the public report suggests greater confidence that non-AI factors explain much of the acceleration than the private report supports. The private report called it plausible that AI-attributable acceleration exceeded a factor of 2 overall by July 15; METR broadly agrees with that assessment while emphasizing that estimating the counterfactual is difficult. METR’s preferred wording says AI models were an increasingly important factor in continuing faster trends, and that overall progress was probably slightly less than twice as fast as without AI assistance, though exceeding 2x remained plausible.

METR also challenges the public report’s treatment of AECI evidence. It says Anthropic omitted private-report qualifications that later snapshots of Model 2 were expected to improve more than the public report suggested; subsequent snapshots improved AECI by about as much as Claude Mythos 5 improved on Mythos Preview. The provided review text ends while METR is disputing Anthropic’s emphasis on weak evidence for a one-time AECI jump.