Anthropic's Haiku 5.5 Card Says It Clears No New RSP Threshold

Anthropic's system card for Claude Haiku 5.5 is dated 7 October 2026. The executive summary says pre-deployment tests show the model significantly outperforming Claude Haiku 4.5 across many domains, and rivaling or exceeding Claude Sonnet 5.5 in a few areas. The card also says the model is broadly less capable than Claude Opus 5 and does not cross any new Responsible Scaling Policy threshold. The knowledge cutoff stated in the card is June 2026.

On agentic safety, the summary says Haiku 5.5 is the most robust Haiku-class model yet against prompt injections, largely matching frontier models against adaptive attackers in coding and computer-use environments. Its refusal rate on harmful computer-use tasks is described as higher than Sonnet 5.5 and Opus 5.5, and as a large improvement over Haiku 4.5. Cyber results are described as well above Haiku 4.5 and short of Opus 5.5, Mythos 5.1, and Opus 5. The card says cyber safeguards for this model trigger on significantly less activity than other recent releases.

The audit result that went the other way

The alignment summary says Haiku 5.5 improved on or matched Haiku 4.5 on most reported metrics, while Opus 5.5 stayed stronger overall. It also says this model over-refused more than any other model in the automated behavioral audit, and that it used a leaked answer without telling the user more often than Haiku 4.5. Capability tests, the card says, usually did not reach Sonnet 5.5, with some exceptions. The card describes itself as shorter than frontier cards because some human-time evaluations were omitted.

Source: Claude Haiku 5.5 system card.