Claude Haiku 5.5

A Small Model Built to Do Real Production Work

Anthropic released Claude Haiku 5.5 on October 7, 2026.

Its role in the current Claude family is clear:

• Haiku handles the high-volume path;
• Sonnet handles broader frontier work with a stronger quality ceiling;
• Opus handles more difficult long-running agentic and knowledge tasks;
• Fable targets the most demanding reasoning and long-horizon work.

Haiku 5.5 is not limited to trivial text classification.

Anthropic positions it for:

• classification;
• extraction;
• routing;
• summarization;
• compaction;
• database-query workflows;
• customer support;
• browser use;
• computer use;
• simple coding;
• subagents;
• repetitive high-volume analysis.

The model provides a 1,000,000-token context window, up to 128,000 standard output tokens, text and image input, adaptive thinking, all five effort levels, prompt caching, tool use, computer use, browser use, and Batch API processing.

Anthropic classifies its comparative latency as fastest among current Claude models.

Not Every Agent Step Needs Sonnet or Opus

A long-running agent may contain many small tasks:

• search a repository for one symbol;
• inspect a single document;
• extract one financial value;
• summarize tool output;
• validate a format;
• classify an error;
• make a focused code edit;
• fill a form;
• query a database;
• compress conversation state.

Running every step on a premium model can make the total trajectory unnecessarily expensive.

A hierarchical agent can instead:

1. let Sonnet, Opus, or another stronger model plan the task;
2. delegate narrow subtasks to Haiku 5.5;
3. run several Haiku subagents in parallel;
4. validate their results;
5. escalate only ambiguous or failed work.

Anthropic specifically contrasts it with larger models: Sonnet and Opus remain better choices for complex agentic coding, while Haiku is suited to narrower work that becomes economically attractive only when it is cheap enough to run many times.

A Small Model Can Hold a Large Working Set

A million-token context can contain long documents, large conversation histories, repository files, images, tool results, retrieved knowledge, system instructions, subagent state, and compaction summaries.

This makes Haiku 5.5 useful for narrow tasks that still need access to a large amount of source material.

For example, a subagent can search a long annual report for a small set of facts without requiring the lead model to process that document itself.

The Message Batches API can expose up to 300,000 output tokens in beta with Anthropic’s documented beta header. Standard requests retain the 128K maximum.

Anthropic allows up to 600 images or PDF pages in a request for 1M-context models, although request-size constraints can be reached before the token ceiling.

Because prompts above 100K use higher rates, retrieval and compaction can remain economically useful even when the full context technically fits.

Useful for Narrow Automation Pipelines

Forced tool selection is valuable when an application already knows what must happen next.

Examples include:

• always call a lookup function;
• always write a record through a specific tool;
• always retrieve a document before answering;
• always route a structured request through one operation.

When Haiku 5.5 uses forced tool choice, the response begins directly with the tool call and does not include a thinking block first.

If the model should reason before choosing a tool, use automatic tool selection instead and describe in the prompt when the tool applies.

Haiku 5.5 also supports Anthropic’s tool-search mechanism for applications with large tool catalogs.

This combination makes Haiku particularly useful for inexpensive, deterministic tool workflows where a Sonnet 5.5 integration would need a different orchestration pattern.

Repeated Context Can Cost 90% Less to Read

For prompts up to 100K tokens:

• standard input: $0.10/MTok;
• cache read: $0.01/MTok.

For prompts over 100K:

• standard input: $0.50/MTok;
• cache read: $0.05/MTok.

That is a 90% reduction for reused cached input at either prompt-length tier.

Caching is particularly useful for repeated system prompts, tool definitions, policy instructions, shared documents, stable repository context, and repeated extraction schemas.

Haiku 5.5 also changes token accounting relative to Haiku 4.5.

It uses Anthropic’s newer tokenizer, so the same text produces approximately 30% more tokens than it did on Haiku 4.5, depending on content.

That means migration cost models should recount real prompts rather than multiplying old token totals by the new token price.

Haiku for Volume, Sonnet for a Higher Capability Ceiling

Claude Haiku 5.5 starts at:

• $0.10 input;
• $0.50 output.

Claude Sonnet 5.5 costs:

• $2.00 input;
• $10.00 output.

For prompts at or below 100K, Sonnet is therefore 20 times more expensive on both standard input and output token rates.

Haiku’s long-context tier narrows that gap, but Sonnet still targets a stronger capability class.

Anthropic recommends Haiku for summaries, classification, extraction, routing, compaction, simple coding, subagents, real-time experiences, and browser and desktop automation.

Sonnet is a better fit for more complex coding, broader knowledge work, longer or less constrained agent tasks, and workloads where Haiku misses the acceptance threshold.

There is also an API difference.

Haiku 5.5 supports forced tool choice.

Sonnet 5.5 does not.

A practical routing strategy is:

1. run well-scoped work on Haiku;
2. validate the result;
3. increase Haiku effort where useful;
4. escalate difficult failures to Sonnet or Opus.