Claude Haiku 5.5 Guide

Claude Haiku 5.5 is Anthropic’s fastest and cheapest model, released on October 7, 2026. It joins Fable 5.1, Opus 5.5, and Sonnet 5.5 in the current Claude lineup.

It is built for:

• High-volume tasks
• Fast responses
• Subagents working under larger models
• Classification and extraction
• Summaries and routing
• Low-cost browser and computer-use agents

Use the smallest model that does the job well. Often, that is now Haiku.

Context window

1 million tokens

Maximum output

128,000 tokens

Effort levels

Low to max, with medium as the default

Thinking

The first Haiku with adaptive thinking

API model ID

claude-haiku-5-5

Prompts up to 100K tokens

$0.10 per million input tokens, $0.50 per million output tokens

Prompts over 100K tokens

$0.50 input and $2.50 output per million tokens

Cache reads

From $0.01 per million tokens

Batch API

50% off

Anthropic says Haiku 5.5 is 75% cheaper on average than Haiku 4.5. One review measured that its tokenizer produces more tokens for the same text, so test your real workload.

Anthropic’s launch benchmarks show a big jump from Haiku 4.5:

OSWorld 2.1

72.4%, up from 15.7% for Haiku 4.5

Terminal-Bench 4.0

39.2%, up from 0.0%

Humanity’s Last Exam with tools

57.4%, up from 18.7%

Sonnet 5.5 still leads Haiku 5.5 on every benchmark Anthropic published. These are vendor results, so test on your own tasks.

• Claude apps, which include Haiku models on every plan
• Claude API
• Amazon Bedrock
• Google Cloud
• Microsoft Foundry

Anthropic has committed not to retire Haiku 5.5 before October 7, 2027.

• Tagging and routing support tickets
• Extracting fields from invoices and forms
• Summarizing long documents
• Moderation queues
• Search and retrieval helpers
• Simple browser automation
• Fast drafts and rewrites

Anthropic positions Haiku 5.5 as a worker under Sonnet 5.5 or Opus 5.5:

• The larger model plans and decides
• Haiku reads files and gathers information
• Haiku handles narrow subtasks in parallel
• The larger model reviews and combines the results

Big model for judgment. Haiku for volume.

Sonnet 5.5

Coding agents and complex multi-step work

Opus 5.5

Hard reasoning and everyday flagship work

Fable 5.1

Long-running agents and the hardest problems

• Change the model ID to claude-haiku-5-5
• Replace manual thinking budgets with the effort setting
• Remove non-default temperature, top_p, and top_k settings
• Remove assistant prefill
• Update computer-use tool versions
• Rerun your evaluations before switching production traffic

Anthropic lists these as breaking changes. Check its migration guide.

• Keep prompts under 100K tokens where possible
• Cache repeated instructions and documents
• Use the Batch API for non-urgent jobs
• Use low or medium effort for simple tasks
• Ask for short, structured outputs

• Weaker than Sonnet 5.5 for coding agents
• Long prompts cost five times more per token
• Safety classifiers can refuse some requests
• It can still make mistakes, so check important outputs

You are a fast assistant for [task]. For each item, return only: [fields] in [format]. If information is missing or unclear, return “unknown” instead of guessing. Keep each answer under [length].