Claude Haiku in one minute
Haiku is a candidate for translation, short summaries, classification and other clearly bounded tasks. This is not shorthand for “cheap and poor”. The trade-off is whether its capabilities meet the task, how quickly a usable result appears and how much the complete processing costs.
Its historical and current role
Haiku became a recognisable family name alongside Sonnet and Opus in Claude 3 in 2024. It emphasised fast, efficient processing. Later versions added capabilities; a historical description should therefore not imply that Haiku can do only what it could at introduction.
Current documentation places Haiku 5.5 in the fastest category at standard speed. Special fast modes, such as those for Opus, are outside that comparison. The model supports text and image input, text output, programming and tool tasks. This makes it more versatile than a simple chatbot, although suitability still depends on task difficulty.
Development progressed through Haiku 3 and 3.5 to Haiku 4.5 on 15 October 2025. On 28 September 2026, Anthropic announced a successor. Haiku 5.5 was released on 7 October 2026 and is now documented as the latest active version. Its introduction highlights high volumes of short tasks, extraction, routing and subagent work alongside Sonnet or Opus.
What can you use it for?
An online shop could classify incoming questions into delivery, returns, product information and other enquiries. Supply examples and separate uncertain cases. The output is a proposal for rules or an employee to check. This is a useful bounded evaluation because the categories and output format are known in advance.
An employee can translate a short text, summarise an email or ask for a simple explanation. In translation, product names, measurements and conditions remain important checks. A fast response adds little value if an essential number changes.
Haiku can also help with focused code tasks or serve as part of an agent. Examples include explaining a limited change or extracting structured fields from text. A clear task contract, such as fixed fields with an “unknown” option, makes usefulness easier to assess objectively.
Speed, cost and capability
A fast model can suit many repetitions, particularly when the task is clear and the answer remains short. Current standard API token prices for Haiku 5.5 are below Sonnet, Opus and Fable. The price list distinguishes prompts up to 100,000 tokens from longer prompts. Total costs also depend on input, output, tools, reasoning and retries.
Assess cost per correctly completed task. If a complicated case requires several failed attempts, another model may ultimately be more efficient. For simple cases, extra capability may contribute little. Examples from your own work reveal the boundary.
A useful setup might route difficult exceptions to Sonnet or Opus. Base that route on checkable signals, such as missing fields or conflicting sources. A model’s stated confidence in its own answer is not a reliable quality measurement by itself.
Context and reasoning in Haiku 5.5
The current API ID is claude-haiku-5-5. Its context window is 1 million tokens. Haiku 5.5 uses adaptive thinking, enabled by default, with an adjustable effort setting and medium as the API default. Model choice and reasoning effort remain different settings.
The earlier Haiku 4.5 has 200,000 tokens of context and uses extended thinking with an explicit token budget. It does not have the same adaptive-thinking and effort configuration. Haiku 5.5 also uses a different tokenizer, which can count the same text differently. Check settings and cost calculations before transferring them from 4.5.
A sufficiently large context window only establishes that information can fit. It does not guarantee correct processing of every detail. For a short classification task, compact, well-structured input may be more appropriate than a complete dossier.
The release and documentation describe access through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Check availability and conditions in the account you use. Release through these channels does not mean every assistant subscription offers the same selection and usage limits.
Which older versions are retired?
Current lifecycle documentation records Haiku 3.5 as retired on 19 February 2026 and Haiku 3 on 20 April 2026. Those specific versions should not be recommended as current choices. Haiku 4.5 is listed as active on 7 October 2026.
For Haiku 4.5, the table also says “not sooner than 15 October 2026”. This is an earliest possible date within the published support commitment, not an announced retirement on 15 October. Without a separate retirement announcement, such a boundary must not be presented as a fixed end date.
Check the channel you use: partner platforms can have separate migration schedules. API availability also does not automatically determine the options visible in a Claude app or subscription.
When should you choose another family?
Test Sonnet for more open-ended writing, code work or iterative feedback. Consider Opus for complex analysis and many dependencies. Fable is relevant to particularly difficult reasoning and sustained processes. Mythos requires specific verification and is not an ordinary upgrade for quick office tasks.
For a first Haiku trial, classify twenty representative questions and assess the results manually. Include some exceptions, record the model version and track correction work. If quality is sufficient, fast processing is a real advantage. Otherwise, adjust the task or model choice based on the errors found.