Claude Haiku 5.5 Official Release EvaluationProvider report
2026-10-07Official release table, System Card summary on page 111 and corresponding sections: 16 metrics and 52 scores. Haiku 5.5 uses adaptive/max; the API default medium is not the evaluation setting. FrontierCode uses the release page's Sonnet 5.5 xhigh score, with max retained separately. OSWorld 2.1 uses the 82-task offline subset; partial and strict are separate from full-set and older offline-version results. HealthBench Professional is length-adjusted. Both Haiku 5.5 Terminal-Bench evaluations use safeguards without fallback. Missing cells create no scores; real zeros are retained. Comparison scores belong to this report without replacing historical snapshots.