Claude Haiku 5.5: 1M tokens, and short prompts ten times cheaper

The small model got big. Anthropic released Claude Haiku 5.5 on 7 October, model ID claude-haiku-5-5, and it changes the math for anything that runs many cheap calls: subagents, summaries, code search, triage.

The price

Up to 100,000 tokens per prompt, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens. Above that it is $0.50 and $2.50. Reading from the cache costs $0.01 per million, writing to the five-minute cache $0.125. Haiku 4.5 was $1 and $5, so short prompts are ten times cheaper, and Anthropic's announcement puts the average saving at about 75%.

One catch eats into that: according to the migration notes, a new tokenizer turns the same text into roughly 30% more tokens. Still a lot cheaper, just not quite as cheap as the price list suggests.

What it can do

• 1M token context and 128K output tokens (Haiku 4.5: 200K and 64K).
• Adaptive thinking with an effort setting, medium by default.
• Available on the Claude API, Amazon Bedrock (anthropic.claude-haiku-5-5), Google Cloud, Microsoft Foundry and Claude Platform on AWS. Anthropic promises not to retire it before 7 October 2027.

Anthropic's own figures against Haiku 4.5 and Sonnet 5.5: Terminal-Bench 4.0 at 39.2% (0.0% and 70.6%), OSWorld 2.1 offline at 72.4% (15.7% and 83.9%), GDPval-AA at 1620 (735 and 1840). Anthropic itself still points to Sonnet 5.5 or Opus 5.5 for complex agentic coding and sees Haiku 5.5 in subagents, summaries and compaction.

What breaks when you switch

The migration notes list the old habits that no longer work:

• budget_tokens for thinking is gone, effort replaces it.
• temperature, top_p and top_k are no longer accepted.
• Prefilling the start of the assistant's answer is not supported.
• The old computer use tool computer_20250124 becomes computer_toolset_20260801.

Our take

Haiku used to be the model you picked when you had to, not when you wanted to. With a 1M window at this price, it becomes the default worker under a bigger model: let Opus or Sonnet plan, let a handful of Haikus read the repo. The breaking changes are a real cost for older code, and the extra tokens from the new tokenizer mean you should measure your own bill instead of trusting the 75%.

In AyCode

AyCode uses Haiku 4.5 today for small side tasks such as chat titles and summaries, and Haiku is not in the chat model menu. Haiku 5.5 is not in the app yet.