Anthropic released Claude Haiku 5.5 on October 7, 2026 for high-volume tasks. Developers can call it as claude-haiku-5-5. Its one-million-token context window allows much longer requests, but the lowest advertised API rate applies only when a prompt has no more than 100,000 tokens.
That boundary matters to teams processing long documents or growing agent histories. Moving from Haiku 4.5 also involves more than changing the model ID: Anthropic’s migration guide identifies request settings and response-handling assumptions that can fail.
The price changes with prompt length
Anthropic’s Haiku 5.5 price table lists these rates in US dollars per million tokens:
For prompts up to 100,000 tokens, Anthropic lists $0.10 per million input tokens and $0.50 per million output tokens.
For prompts over 100,000 tokens, the listed rates are $0.50 per million input tokens and $2.50 per million output tokens.
The listed input and output rates are five times higher in the longer-prompt tier. Cache reads and writes have their own two-tier rates. The threshold concerns prompt length, while the applicable tier also determines the listed output rate. A one-million-token context window is therefore a capacity limit, not a promise of the lowest rate throughout that window.
Anthropic says about 90% of requests to Haiku 4.5 were within 100,000 tokens. That describes Anthropic’s previous-model traffic, not the prompt distribution of a particular application. Teams should measure their own requests, especially those near the threshold.
The same text can consume more tokens
Haiku 5.5 uses a newer tokenizer. Anthropic says the same input text produces approximately 30% more tokens than on Haiku 4.5, although the increase depends on the content. That can move a request across the price boundary and alter estimates built from old token counts. It can also make an existing max_tokens value too small for a comparable response.
Adaptive thinking is on by default. Applications that send Haiku 4.5’s thinking: {"type": "enabled", "budget_tokens": N} must change that setting; Anthropic says the old form returns a 400 error on Haiku 5.5. Clients should also select response content blocks by type rather than assuming the first block contains the answer, because it may contain thinking.
Other compatibility checks are concrete: Anthropic instructs developers to remove temperature, top_p and top_k settings, and to replace final assistant-message prefills with another approach. Its guide says incompatible values or an assistant prefill can return a 400 error. Integrations using the older computer-use tool on the Claude API or Google Cloud need the newer toolset.
Test the workload, not just the model ID
A migration test should count tokens with Haiku 5.5, sample requests on both sides of the 100,000-token boundary, and compare cost per completed task rather than relying on a per-token headline rate. It should also exercise parsers, thinking-block handling, output limits and requests that used the deprecated settings. Organizations with a Haiku 4.5 Priority Tier commitment need a separate capacity plan because Anthropic says that tier is not supported for Haiku 5.5.
Anthropic lists Haiku 5.5 as available through its API and major cloud platforms. Teams buying through a cloud provider should verify that provider’s actual billing and availability before projecting costs. Anthropic’s published benchmarks and speed claims were not independently tested for this article.
