CONFIRMED: a small model changes the automation budget
Anthropic launched Claude Haiku 5.5 on October 7 with a concrete pitch: lower the cost of the short, repetitive tasks that support business agents and applications. On its application programming interface, or API, pricing starts at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. A token is the unit into which a model divides and counts text. The reduction against Haiku 4.5 is significant, but one condition matters: above that threshold, prices rise to $0.50 and $2.50 respectively.
The work Haiku is meant to take on
The release targets classification, data extraction, summaries, request routing and subagents, specialized workers called by a lead agent. For a Brazilian small or medium business, this could mean categorizing support messages, organizing catalog fields or summarizing documents before a more expensive model makes a decision. These are use case examples, not results measured by MaxAssistant. The economic change is being able to test this high volume layer without assigning every task to the most powerful model. That makes model selection an operational design choice rather than a prestige contest.
More context does not mean low pricing at every size
Haiku 5.5 has a 1 million token context window and supports up to 128,000 output tokens. More material can fit in a conversation, but that does not remove the higher pricing tier. The documentation also warns that its newer tokenizer counts approximately 30% more tokens for the same text than Haiku 4.5. In an independent check, Simon Willison found around 25% more tokens for one long prompt. The increase varies with content; both observations explain why businesses should recalculate their budgets using actual documents rather than historical token counts.
Adjustable reasoning requires a real migration
The model brings adaptive thinking to the Haiku family, enabled by default, with an effort setting to balance quality, speed and consumption. Updating an old integration takes more than changing a model name. Anthropic documents incompatibilities with manual thinking budgets, nondefault sampling parameters and assistant message prefills. Applications must also identify response blocks by type: the first block may contain thinking rather than the final answer. Browser use support is part of the offering, but the ability to operate a screen does not remove the need for permissions, bounded access and result verification.
Cheaper cache reads and subscriber API credits
Anthropic also reduced Sonnet 5.5 cache read pricing from $0.20 to $0.10 per million tokens. A cache reuses previously processed parts of a conversation; this is not a discount on every token. The company estimates roughly 20% lower costs on much agentic work, depending on usage. Monthly API credits are rolling out this week to Max and Team subscribers: $100 for Max 5x, $200 for Max 20x and up to $500 pooled for Team. Willison verified the option to link an API organization. This is not unlimited API access, and unused credits do not roll over.
Cheap for the right task, not a universal replacement
The model is announced for the Claude Platform and Amazon, Google and Microsoft clouds. Claude Code 2.1.293 also adds Haiku 5.5 as the default Haiku model on Anthropic’s API. The company itself still recommends Sonnet and Opus for complex agentic coding. Our assessment: test Haiku on bounded tasks, comparing accuracy, latency, human review and cost per completed result. Lower token prices can make an operation viable. Excessive context, high reasoning effort and repeated attempts can consume the advantage, so the meaningful comparison is the finished job, not the headline rate.
Sources
Anthropic: Claude Haiku 5.5 announcement; Anthropic: pricing and specifications; Anthropic: changes and migration; Simon Willison: independent test and pricing analysis; Anthropic: monthly API credits; Anthropic: Claude Code 2.1.293.