Claude Haiku 5.5 lands: 90% cheaper per token, way smarter

Anthropic just made its small model much cheaper. Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens on requests under 100K tokens. That's 90% below Haiku 4.5's list price. Above 100K tokens it's $0.50/$2.50, a 50% cut. Once you account for request sizes and the new tokenizer, Anthropic says it costs about 75% less to run on average. That puts it level with OpenAI's GPT-6 Luna.

The benchmark gains are even bigger than the price cut. On OSWorld 2.1 (computer use) it scores 72.4%, up from 15.7% for Haiku 4.5. On GDPval-AA v2.1 (knowledge work) it scores 1,620, up from 735. With tools, it gets 57.4% on Humanity's Last Exam, up from 18.7%. It's also the first Haiku with an adjustable effort setting, and that setting changes the results a lot. On Terminal-Bench 4.0 (agentic coding) it scores 39.2% at max effort and about 20% at the default medium setting.

It's live now on the Claude Platform, AWS, Google Cloud and Microsoft Azure as claude-haiku-5-5. Anthropic is aiming it at high-volume work like summaries, classification, database queries and context compaction.

Why it matters: a cheap small model can now actually operate a computer, which changes the math for anyone running agents at scale.