The model costs $0.10 per 1M input tokens and $0.50 per 1M output tokens for requests under 100k tokens, which is 90% less than Haiku 4.5. However, rates increase fivefold to $0.50 and $2.50 for requests exceeding that limit. This new iteration introduces adjustable effort settings and has already outperformed GPT-6 Luna on OSWorld computer use, scoring 72.4%, and Terminal-Bench agentic coding, where it reached 39.2%.
Tuesday, October 20, 2026
Story
Claude Haiku 5.5 Ranks 3rd on Vibe Code Bench
AI13 days ago
Anthropic's Haiku 5.5 achieved a score of 90.4% on the Vibe Code Bench, climbing 89 positions from its predecessor to become the third most capable model on the leaderboard. Despite this performance gain, the model is more expensive to run per test than Haiku 4.5 across every shared benchmark due to higher token consumption. On legal research tasks, for example, Haiku 5.5 averaged 59 steps and wrote approximately 15 times as much as the previous version, with 80% of the output dedicated to reasoning.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.