Anthropic releases Claude Opus 5.5 with system card detailing RSP safety evaluations

Anthropic released Claude Opus 5.5 alongside a system card covering its Responsible Scaling Policy (RSP) evaluations: safety classifiers, bio/cyber risk tests, AI R&D autonomy checks and agentic safety testing. The card reports on-trend capability gains, not a threshold crossing: the model stays at CB-1 (not CB-2) and Autonomy-1 (not Autonomy-2). One real policy change is buried in the detail: Anthropic has stopped testing helpful-only model versions for bio risk and switched to refusal-avoidant test design instead.

This is a routine capability update dressed up as news, useful mainly to people who track RSP compliance and model safety testing methodology rather than general audiences. The bio-eval testing change is the one item worth attention: it alters how Anthropic checks whether a model can be coaxed into providing dangerous biological information, which matters to biosecurity researchers, red-teamers and anyone relying on Anthropic's safety claims. For developers and enterprises, the practical takeaway is a marginally better, cheaper model with no new risk classification, so no urgent action is required beyond normal model-swap evaluation.

The "world's most powerful model" framing rests on selected benchmarks (Artificial Analysis) and a comparison against a rival model referred to inconsistently as "Fable 5.1" and "Mythos 5.1" in the source material, which suggests the underlying commentary itself has extraction or naming errors and should not be taken as a precise comparison. There are no primary links in this extraction: the analysis is a LessWrong write-up of the card, not the card itself, so specific numbers and claims have not been independently verified here. Calling this a safety milestone is overstated; the card itself says capabilities are on-trend and no risk threshold was crossed.

1. 01Read the actual Anthropic system card (not secondary commentary) to verify the bio-eval methodology change and confirm CB-1/Autonomy-1 classifications with primary source numbers
2. 02Check whether independent evaluators (e.g. CAISI, UK AISI) publish their own assessment of Opus 5.5 that confirms or contradicts Anthropic's self-reported RSP results
3. 03Watch for the next model release (from Anthropic or the unnamed rival) to see if the bio-eval testing change becomes an industry norm or stays a one-off Anthropic policy shift