A lab reports its AI R&D evaluations have saturated

Anthropic’s redacted August 2026 Risk Report, published alongside the disclosure of an unreleased Model 2 stronger than Mythos 5, states that the lab’s AI R&D evaluations have saturated, with Claude now authoring most of the code merged into Anthropic’s own production repositories, while the lab moves its misalignment-risk rating from “very low” to “low” (risk report; Axios). Watchers did the arithmetic on the redacted numbers: Model 2 beat Mythos 5 by 12.5 points on CoBench v2, whose 85% threshold marks researcher replacement, so “2027 is the takeoff” (X). In the trajectory this is the point where the measuring instrument gives out: it follows the lab’s own When AI builds itself evidence and the March figure of 70–90% of model code, extends the effectively all product code claim to the repositories behind the models, and is answered in the same issue by the Conceptual Reasoning Index, a new instrument for what the old ones can no longer score.