Claude Haiku 5.5

Claude Haiku 5.5 (released Oct 7, 2026) completes the Claude 5.5 family as its smallest member. It introduces two-tier pricing: prompts up to 100K tokens cost $0.10/$0.50 per 1M input/output tokens (matching OpenAI's GPT-6 Luna), stepping to $0.50/$2.50 above that. Anthropic says it averages ~75% cheaper to run than Haiku 4.5. It is the first Haiku-class model with an adjustable effort setting (adaptive thinking, default medium), a 1M-token context window, 128K max output, and text+image input. Quality leapt on agentic small-model tasks: Terminal-Bench 4.0 went from 0.0% (Haiku 4.5) to 39.2%, OSWorld 2.1 to 72.4%. It became generally available in GitHub Copilot and on Vercel AI Gateway within a day of launch. Caveat: a new tokenizer uses ~30% more tokens per text than Haiku 4.5's, so token-level pricing understates real cost somewhat.

Provider-reported (Anthropic, system card published): GDPval-AA v2.1 1620 (Haiku 4.5: 735; GPT-6 Luna: 1437); OSWorld 2.1 72.4% offline subset (4.5: 15.7%; Luna: 48.9%); Terminal-Bench 4.0 39.2% (4.5: 0.0%; Luna: 16.4%); FrontierCode 1.1 46.4% (Luna: 42.4%); Humanity's Last Exam 57.4% with tools (4.5: 18.7%); Chartography 46.4% (4.5: 6.4%; Luna: 29.1%).

Independent: Artificial Analysis Intelligence Index gives Haiku 5.5 a 43 at max effort — the top of the small-model group (agentbreaking, Oct 8, 2026). Independent reviews (agentbreaking, developersdigest, Neowin, datanorth) confirm specs and pricing; Neowin flagged the ~30% higher token counts from the new tokenizer as a real-world cost caveat. Asana reported >30% lower latency vs its previous model in its own evals (Anthropic-hosted quote).