Claude Haiku 5.5 is here: a much smarter small model at a tenth of the price

Anthropic released Claude Haiku 5.5 on October 7, 2026. It is the third and final model in the Claude 5.5 family, after Opus 5.5 on September 22 and Sonnet 5.5 on September 28. Anthropic calls it its cheapest, fastest and most capable small model yet, and the numbers back that up: it is far smarter than Haiku 4.5 and costs a fraction as much.

What is new in Haiku 5.5

Haiku 5.5 is built for work where volume and speed matter more than depth: classification, extraction, routing, summaries, conversation compaction and database queries. It is also designed to work as a subagent, handling narrow jobs for a larger Opus 5.5 or Sonnet 5.5 model that plans the overall task. Anthropic also points to live customer support and browser automation as good fits.

• Effort control. Haiku 5.5 is the first Haiku model with an adjustable effort setting: low, medium, high, xhigh or max. The default is medium.
• Adaptive thinking. The model decides when and how much to think, guided by effort. At low effort it can skip thinking on simple requests.
• Big context, big output. A 1 million token context window and up to 128K output tokens, the same as the rest of the 5.5 family.
• The fastest current Claude model. Anthropic lists it as the fastest model in the lineup at standard speed. (Opus models in fast mode can still generate output more quickly.)
• Fresher knowledge. Its training data runs to June 2026.

Pricing: the headline feature

Haiku 5.5 has two price tiers, depending on how long your prompt is:

For short prompts, that is a 90% cut on every line. For prompts over 100,000 tokens, it is a 50% cut. Anthropic says about 90% of the requests sent to Haiku 4.5 fall under the 100K line, so most users land in the cheap tier. The Batch API takes a further 50% off.

Anthropic also cut a price on the side. Sonnet 5.5 cache reads dropped from $0.20 to $0.10 per million tokens on the same day, which Anthropic says makes most agentic work on Sonnet 5.5 about 20% cheaper.

Benchmarks

Anthropic compared Haiku 5.5 with Haiku 4.5, OpenAI's GPT-6 Luna (which has the same per-token prices) and Sonnet 5.5:

Bold marks the best score among the small models. Haiku 5.5 beats GPT-6 Luna on every test where both have a score, often by a wide margin: 72.4% vs 48.9% on OSWorld, and 39.2% vs 16.4% on Terminal-Bench. The gap to Haiku 4.5 is even bigger.

Sonnet 5.5 is still clearly ahead, though, especially on agentic coding (70.6% vs 39.2% on Terminal-Bench 4.0). Anthropic itself recommends Sonnet 5.5 or Opus 5.5 for complex coding work, and Haiku 5.5 for narrow, well-defined jobs.

What early customers say

Anthropic shared results from companies that tested Haiku 5.5 before launch:

• Asana saw task completion latency drop by more than 30%, and up to 2.5× faster inference per agent turn, compared with the model it uses today.
• HubSpot got a 92.8% score on its CRM test suite, the best it has seen there.
• AlphaSense measured 0.84 vs 0.76 for Haiku 4.5 on its "Ask in Document" feature, which handles about 8 million calls a week.
• Box saw scores 11 points higher than Haiku 4.5 at about half the latency.
• Cognition paired Haiku 5.5 as a sidekick to Opus 5.5 in Devin, and says the pair cut cost and latency while scoring 66.2 on FrontierCode.

Safety

Anthropic says Haiku 5.5 is a major step up from Haiku 4.5 on nearly every alignment evaluation, with far fewer cases of misaligned behaviour and less willingness to cooperate with misuse.

Its cybersecurity safeguards are stricter than Haiku 4.5's, but looser than on other recent models: they allow a wider range of defensive security work than Sonnet 5.5, while still blocking penetration testing and attacker-style techniques. Its biology safeguards match Sonnet 5.5 and Opus 5. Researchers who need broader access can apply to Anthropic's Cyber and Life Sciences verification programs.

Where you can use it

• Claude API: model ID claude-haiku-5-5. It is a fixed ID with no date suffix.
• Cloud platforms: Amazon Bedrock (anthropic.claude-haiku-5-5), Claude Platform on AWS, Google Cloud and Microsoft Foundry (all claude-haiku-5-5).

Haiku 4.5 is still available, so you can move over at your own pace.

For developers: what changes when you upgrade

Haiku 5.5 is a bigger jump from Haiku 4.5 than the version number suggests. Code written for Haiku 4.5 can break, so read the migration guide before swapping the model ID.

A few more things to watch for:

• Recount your tokens. The new tokenizer means the same text uses about 30% more tokens, and large images can cost about 2.5× as many tokens. Revisit max_tokens limits and cost estimates.
• Thinking is on by default. A response can start with a thinking block even if you did not ask for one, so pick content blocks by type, not by position. Thinking tokens also count toward max_tokens.
• Thinking text is hidden by default. Thinking blocks come back empty. Set display: "summarized" if you want to see a summary.
• Forced tool use still works, unlike on Opus 5.5, but the model then skips thinking. Use tool_choice: { type: "auto" } if you want it to reason first.
• On Amazon Bedrock, structured outputs are not available for Haiku 5.5.

Here is a classification call that sets effort on purpose:

import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic();

const message = await client.messages.create({
  model: 'claude-haiku-5-5',
  max_tokens: 4000,
  // No temperature, top_p or top_k: Haiku 5.5 rejects them.
  output_config: { effort: 'low' }, // low | medium | high | xhigh | max
  messages: [
    {
      role: 'user',
      content: 'Classify this support ticket as billing, bug or feature request: "I was charged twice this month."',
    },
  ],
});

// Thinking blocks can come first, so pick the text block by type.
for (const block of message.content) {
  if (block.type === 'text') console.log(block.text);
}

For simple, high-volume tasks like this one, low effort keeps both cost and latency down. Raise it for harder work.

Should you switch?

For most Haiku users, yes. You get a far more capable model for a fraction of the price. That is a rare combination.

• Running Haiku 4.5 at scale? Switch. Even after the tokenizer change, Anthropic expects your bill to drop by about three quarters. Just fix the breaking changes first.
• Using Sonnet for simple, high-volume tasks? Test Haiku 5.5. On knowledge work and computer use it is now close enough to Sonnet 5.5 that the 20× lower price may win.
• Building agents? Try Haiku 5.5 as a subagent for narrow jobs like search, extraction and summaries, with Opus 5.5 or Sonnet 5.5 in charge.
• Doing complex coding? Stay on Sonnet 5.5 or Opus 5.5. That is Anthropic's own advice.

With Haiku 5.5 out, the Claude 5.5 family is complete: Opus for the hardest work, Sonnet for the best mix of speed and smarts, and Haiku for everything that needs to be fast and cheap.