What overall risk rating will Anthropic assign to the automated AI R&D threat model in the next Risk Report it publishes after 2026-08-28?

rationale:

Anthropic's most likely move in its next Risk Report is to maintain its "Low" rating (48%) for the automated R&D threat model. The Claude Opus 5.5 system card published in late September explicitly noted that the threshold is not currently met https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf. Opus 5.5 scored 55.8% on CoBench 2.1, well below the 85% mark Anthropic associates with full substitution https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf. Furthermore, internal measurements do not yet show a sustained 2x acceleration in overall AI progress, with third-party evaluator METR estimating roughly a 1.5x acceleration https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf How an assessor could investigate evidence of .... Because Anthropic appears to tie its overall rating closely to whether the hard RSP criteria are met, "Low" remains the modal outcome.

However, there is a substantial chance (36%) that Anthropic upgrades the rating to "Moderate or higher" without declaring the threshold met. In the August 2026 report, Anthropic noted declining confidence in its "Low" rating due to saturating evaluations and explicitly warned that the threat model could become a "major concern in the next 6-12 months" https://www.anthropic.com/aug-2026-risk-report https://www.anthropic.com/aug-2026-risk-report. Since then, Anthropic reported that Claude now "leads" (AL4) 26% of its internal AI R&D work—up from under 1% in February 2026 Measurements for understanding the pace of AI ... Measurements for understanding the pace of AI .... Given this trajectory, the timeline implied by related estimates of threshold crossings, and the fact that "Moderate" is a label already used in Anthropic's sub-assessments https://www.anthropic.com/aug-2026-risk-report, an interim step up is highly plausible as a conservative measure before the threshold is officially crossed.

While less likely, the probability of Anthropic officially stating the threshold has been met (11%) within the next report's coverage window is a meaningful tail risk. Even though the Opus 5.5 system card suggests the substitution and acceleration benchmarks are still out of reach, METR assigned a roughly 30% chance to a 2x acceleration having already occurred https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf. Additionally, CEO Dario Amodei recently noted that progress has been "drastically faster" since summer 2026 https://darioamodei.com/post/we-must-pace-the-frontier. The recent embedding of six METR staff members with "employee-like access" and the right to publish without an Anthropic veto https://www.anthropic.com/news/an-update-on-independent-evaluators https://www.anthropic.com/news/an-update-on-independent-evaluators adds a layer of independent scrutiny that makes it harder to understate the risk if internal measurements start breaching the 2x slope change requirement.

The extreme outcomes are highly unlikely. Reverting to "Very low" (2%) would require reversing a clear, documented trend of accelerating AI involvement in R&D and contradicting Anthropic's own warnings about saturated evaluations. Finally, there is only a 3% chance that no Risk Report is published before April 30, 2027. Under RSP v3.4, a report is required every 3-6 months Anthropic's Responsible Scaling Policy Responsible Scaling Policy, placing the deadline for the next one between mid-November 2026 and mid-February 2027. Missing the late-April cutoff would require Anthropic to delay publication by more than two months beyond its own policy commitment, which is improbable barring a major, unannounced restructuring of the RSP.