Haiku 5.5 is cheap at $0.10 per million input tokens, but only for prompts up to 100,000 tokens. Above that line every rate jumps fivefold, adaptive thinking is on by default at medium effort, and thinking tokens count against max_tokens. Size your prompts and limits before you migrate.
Anthropic released Claude Haiku 5.5 on October 7, and the headline price is what most teams noticed first: $0.10 per million input tokens and $0.50 per million output tokens. Read the pricing table closely, though, because those rates apply only to prompts up to 100,000 tokens. Send a prompt over 100,000 tokens and Haiku 5.5 bills $0.50 per million input tokens and $2.50 per million output tokens, five times the headline rate on both sides. Cache reads follow the same step, from $0.01 to $0.05 per million tokens.
That threshold matters more than it looks, because Haiku 5.5 also has a 1 million token context window, up from 200,000 on Claude Haiku 4.5. The model invites you to stuff in whole document sets and long histories, and the moment a request crosses 100,000 tokens you are paying a very different price. The newer tokenizer adds to the surprise: Anthropic says the same text counts as approximately 30 percent more tokens than on Haiku 4.5, so a prompt that sat comfortably under the line on the old model can land over it on the new one.
The second trap is thinking. Adaptive thinking is on by default, and the default effort on the Claude API is medium, so Haiku 5.5 may think before it answers even when your request never mentions thinking. Those thinking tokens count toward max_tokens. A tight limit that worked for a classifier on Haiku 4.5 can now stop after a thinking block and before any text, which looks like an empty answer to code that reads the first content block. Thinking text is also omitted by default; set thinking.display to summarized if you want to see it.
A few old habits now fail outright. A non default temperature, top_p or top_k returns a 400 error, and manual extended thinking with budget_tokens returns an error too. Steer the model with the effort parameter instead. For high volume jobs that can wait, the Batch API takes 50 percent off input and output.
Try it today
- Recount your real prompts with the Haiku 5.5 tokenizer and flag any that land over 100,000 tokens.
- Split, trim or cache long contexts so routine requests stay under the 100,000 token pricing line.
- Raise max_tokens to leave room for thinking, or lower effort for simple classification and routing.
- Select response content blocks by type, not position, so a leading thinking block never reads as the answer.
- Remove temperature, top_p, top_k and budget_tokens from requests before switching model IDs.
- Move overnight and bulk workloads to the Batch API for the 50 percent discount.