Anthropic released Claude Haiku 5.5 on October 7. The part that matters to me is the cost of all the small jobs: lookups, summaries, classification, and subagents helping a larger model.
For prompts up to 100,000 tokens, input is $0.10 and output is $0.50 per million tokens. That’s a 90% price cut from Haiku 4.5. Above that threshold, it’s $0.50 and $2.50. Haiku also gets adjustable effort for the first time, so you can trade speed and cost for more reasoning.
That higher price tier matters for agents: their context is sent again on each turn. Keeping workers focused and their conversations short helps keep the small jobs cheap.
Two results from Anthropic’s release benchmarks:
- OSWorld 2.1, the offline computer-use subset: 72.4%, up from Haiku 4.5’s 15.7%.
- Terminal-Bench 4.0, for work in a terminal: 39.2%, compared with Sonnet 5.5’s 70.6%.
Those are Anthropic’s numbers. The terminal result is a useful reminder to keep the harder coding work on a stronger model.
I’ve updated my agents and skills repos to account for Haiku 5.5, including routing, effort settings, and context guidance. The API and migration details are in there too. Check your provider’s model aliases when switching: haiku can still resolve to 4.5 outside Anthropic.