Anthropic releases Claude Haiku 5.5 System Card

Anthropic releases the System Card for Claude Haiku 5.5, reporting pre-deployment evaluation results across seven areas: RSP, cybersecurity, guardrails and helpfulness, agent safety, alignment, model welfare, and capabilities. RSP evaluations show Haiku 5.5's overall capabilities are below Claude Opus 5, and it does not cross the CB-2 or Autonomy-2 thresholds, but it does meet the CB-1 and Autonomy-1 thresholds and applies the corresponding mitigations. On cyber capabilities, it is notably stronger than Haiku 4.5 but weaker than Opus 5.5, Claude Mythos 5.1, and even Opus 5, so its cyber guardrails trigger on a narrower range than other recent releases. Agent safety evaluations call it the most robust Haiku-tier model against prompt injection to date, with robustness approaching frontier models against adaptive attackers in coding and computer-use environments. The model's knowledge cutoff is June 2026.