Claude Haiku 5.5 costs 10 cents per million tokens
Key takeaways
- Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for requests under 100,000 tokens.
- Anthropic's pricing is roughly 75 percent below Haiku 4.5 on average, and up to 90 percent lower on common short prompts.
- It adds a 1 million token context window and an adjustable effort setting.
Last updated: 9 October 2026
Ten cents buys a million input tokens on Claude Haiku 5.5, which Anthropic released on 7 October. Output costs $0.50 per million, both for requests under 100,000 tokens.
Anthropic's pricing works out roughly 75 percent below Haiku 4.5 on average, and up to 90 percent cheaper on the common short prompt range. The model also adds a 1 million token context window, an adjustable effort setting, and stronger computer use and agent benchmark scores than its predecessor.
What Claude Haiku 5.5 pricing looks like in practice
A token is a chunk of text, roughly three quarters of a word. Say an agent reads 200,000 tokens and writes 20,000 on each run. On Haiku 5.5 that costs $0.02 for input plus $0.01 for output, so $0.03 a run. A hundred runs come to $3.
The same job on Sonnet 5.5, priced at $2 in and $10 out, costs $0.40 plus $0.20, or $0.60 a run. Our piece on Claude Sonnet 5.5 pricing has the details of that tier.
Why it matters
When a small model costs this little, running an agent all day stops being a budget decision. Tasks that loop, such as checking a mailbox or watching a log, make many calls, and per call cost is what decides whether they are affordable. Our explainer on how AI agents work shows why those loops add up.
Anthropic also halved Sonnet 5.5 cache-read prices and began issuing monthly API credits to Max and Team subscribers, so the savings reach beyond the smallest model. If you run assistants on your own server, the guide to OpenClaw, the self-hosted AI assistant is a good place to see where a cheap model fits.
What to watch
Price per token is not price per task. A cheaper model that needs more retries can cost more in the end, so wait for independent tests on real agent jobs before treating the 75 percent figure as your saving. The next signal is whether rival labs answer with their own cuts to small model prices.
