According to Anthropic, Claude Haiku 5.5 is its cheapest, fastest and most capable small model, built for high-volume work such as summarisation, compaction, database queries and classification, and intended to serve as a subagent alongside Opus 5.5 and Sonnet 5.5. The company says it is the first Haiku-class model with an adjustable effort setting, running from Low to Max, and that it is its fastest model at standard speed.
On the evaluations Anthropic published, Haiku 5.5 scores 72.4% on the offline subset of OSWorld 2.1, against 15.7% for Haiku 4.5 and 83.9% for Sonnet 5.5; 39.2% on Terminal-Bench 4.0, against 0.0% for Haiku 4.5 and 70.6% for Sonnet 5.5; 45.9% on Humanity's Last Exam without tools and 57.4% with them; and 1620 on GDPval-AA v2.1, against 735 for Haiku 4.5 and 1840 for Sonnet 5.5.
Anthropic prices Haiku 5.5 about 90% below Haiku 4.5 for requests up to 100,000 tokens and 50% below above that, which it says works out to roughly 75% lower cost per task on average; it notes the new tokenizer uses slightly more tokens for the same work. Alongside the launch it halved cache-read pricing for Sonnet 5.5 and added computer-use and browser-use support to its Python and TypeScript SDKs in beta.
Our takeThe OSWorld jump is the number to scrutinise: both figures come from the offline subset, and Anthropic's own effort-setting charts show a wide spread across Low through Max, so scores are only comparable at matched settings. Note where the small model still loses — 39.2% on Terminal-Bench against Sonnet's 70.6% — long-horizon agentic coding remains the gap an effort dial does not close. For developers the practical question is which effort setting each published number was measured at.