Claude Haiku 5.5 Scores 72.4% on OSWorld and 1620 on GDPval-AA — at Roughly a Quarter of Haiku 4.5's Running Cost
GA Oct 7 on the Claude Platform, AWS, Google Cloud and Azure. Anthropic-reported: GDPval-AA v2.1 1620 (Haiku 4.5: 735; Sonnet 5.5: 1840); OSWorld 2.1 offline 72.4%; HLE 45.9% no tools; Terminal-Bench 4.0 39.2% (Sonnet 5.5: 70.6%).
Anthropic released Claude Haiku 5.5 on October 7, its small, fast model for high-volume work: summaries, classification, database queries, subagents. It's the first Haiku with an adjustable effort setting, and Anthropic says it costs about 75% less to run than Haiku 4.5.
Benchmarks (Anthropic launch post)
- GDPval-AA v2.1 (knowledge work): 1620 (Haiku 4.5: 735; GPT-6 Luna: 1437; Sonnet 5.5: 1840)
- AA-Briefcase v1.1: 1578 (Haiku 4.5: 614; Sonnet 5.5: 1824)
- OSWorld 2.1, offline subset (computer use): 72.4% (Haiku 4.5: 15.7%; Sonnet 5.5: 83.9%)
- Humanity's Last Exam: 45.9% no tools / 57.4% with tools (Sonnet 5.5: 56.9% / 64.5%)
- Terminal-Bench 4.0: 39.2% (Sonnet 5.5: 70.6%)
- FrontierCode 1.1 (Main): 46.4%; Chartography: 46.4% no tools
Price
Per million tokens, prompts up to 100k: $0.10 input, $0.50 output (Haiku 4.5: $1 / $5). Over 100k: $0.50 / $2.50. Sonnet 5.5 cache reads were also halved to $0.10, which Anthropic says cuts its agentic cost by about 20%.
Read it straight
All scores are self-reported by Anthropic. The coding gap to Sonnet 5.5 is still wide, and Anthropic says so. The tokenizer also changed, so the same task uses slightly more tokens.
Sources: Anthropic — Introducing Claude Haiku 5.5.
