Anthropic released Claude Haiku 5.5 on October 7. The small AI model is designed to handle recurring tasks at lower cost, such as sorting messages, extracting information from documents, or helping a larger agent. For developers, the performance improvement is only part of the story: A new pricing structure and a different way of breaking down text determine how much an application actually saves.
Key takeaways
- Haiku 5.5 is launching on the Claude Platform and through major cloud partners. Its direct API model ID is claude-haiku-5-5.
- For prompts of up to and including 100,000 input tokens, a million input tokens cost $0.10 and a million output tokens cost $0.50. Longer prompts incur higher rates.
- The new tokenizer counts the same text differently. Lower token prices therefore do not translate directly into equivalent savings per job.
- Adjustable thinking effort makes Haiku more flexible. Anthropic still recommends larger models for demanding coding agents.
Where a small model can take on substantial work
Haiku’s most interesting role is handling bounded tasks with a clear answer format. A support system, for example, needs to determine whether a message concerns an invoice, a cancellation, or a technical problem. A document assistant might extract individual facts from a report. These steps do not automatically require the same model that handles a lengthy research assignment or a difficult software change.
This opens up a useful division of labor: A larger model plans, while a smaller one handles specific subtasks. Such a helper, often called a subagent, receives only the material and instructions it needs. Our assessment: When these jobs happen frequently, a cheaper model can matter more than another top score on an occasional specialized task. The prerequisite is that its results are reliable enough for the intended purpose.
Anthropic’s published benchmarks show a substantial improvement over Haiku 4.5. In the OSWorld 2.1 computer-use test, the company reports 72.4 percent versus 15.7 percent, specifically on the offline subset. In the Terminal-Bench 4.0 coding-agent test, Haiku 5.5 scores 39.2 percent compared with Sonnet 5.5’s 70.6 percent. These are vendor results from defined test environments, rather than success rates for your own workplace.
The gap nevertheless helps explain why switching by task type makes more sense than replacing a model everywhere. Someone delegating multistep software work has different requirements from someone sorting incoming text into fixed categories. The larger Claude Opus 5.5 remains a different class of tool. Haiku adds a cheaper candidate for the many small steps in between.
What the new pricing structure actually means
The official API prices distinguish between prompts of up to and including 100,000 tokens and longer inputs. Tokens are the chunks of text the model processes; they do not simply correspond to words. In the cheaper tier, a million input tokens cost $0.10 and a million output tokens cost $0.50. Above the threshold, those rates become $0.50 and $2.50, respectively. Input length therefore also determines the price of the answer.
Consider a simple calculation: A job with 10,000 billed input tokens and 1,000 billed output tokens costs $0.0015 without caching or additional fees. A job with 120,000 input tokens and the same output volume costs $0.0625. This is not an application measurement. It illustrates why the cheapest pricing row is insufficient for estimating the cost of processing large document collections.
The tokenizer, the mechanism that breaks text into tokens, adds another consideration. According to the model documentation, the same text counts as approximately 30 percent more tokens on Haiku 5.5 than on Haiku 4.5; the exact difference depends on the content. This implies neither a fixed cost increase nor a fixed saving. Anyone migrating needs to recount real inputs with the new model and account for the answers it actually produces.
Caching also deserves separate attention. Prompt caching can reuse recurring input sections at lower cost, such as fixed instructions or parts of documents. The documentation requires matching prompt sections to be identical. A changing opening or constantly modified tool descriptions can undermine the expected benefit. A sound estimate therefore combines input, output, cache hits, and repeated attempts after incorrect answers.
How to evaluate a migration
On the direct Claude API, an experiment starts with the model ID claude-haiku-5-5. An existing project should not blindly replace that string alone, however: The migration guide lists changes involving request parameters, handling thinking blocks, and existing conversation histories. This matters to developers because a cheaper request is of little use if the application subsequently reads its answer incorrectly.
Adjustable thinking effort is new to the Haiku line. The documentation recommends medium as a starting point. For short, simple tasks, low can be faster and cheaper; on longer agent assignments, however, the model may then be more likely to skip a search or a check. More thinking effort is therefore a control to tune, rather than a free improvement. It should be compared using precisely the jobs the application will later pay to process.
A useful comparison starts with a small collection of real cases: ordinary messages, incomplete information, and difficult edge cases. Both models can then be checked for correct categories, extracted values, and output formats. Processing time and cost per usable result also belong in the evaluation. A model that needs more rework can be the worse choice despite its lower list price.
The opportunity is more affordable steps
Haiku 5.5 primarily makes an architecture more attractive in which every request does not go to the largest model. The practical advance would be an assistant that can look things up, sort material, and prepare work more frequently without making every small step expensive. The size of that gain depends on your own jobs. The compelling use case is a clearly bounded step that works reliably and makes the overall process faster or cheaper.