What Is Claude Haiku 5.5? Pricing, Performance, How to Choose It Over Sonnet 5.5, and What to Watch When Switching From Haiku 4.5

Claude Haiku 5.5 (model ID claude-haiku-5-5) is the lightweight Claude model that Anthropic released on October 7, 2026. It is built for high-volume, speed-sensitive work such as classification, extraction, routing, and subagents. Its price is one-tenth that of the earlier Haiku 4.5 (for prompts up to 100,000 tokens), and its context window has expanded to 1M tokens. On the other hand, code written for Haiku 4.5 includes some patterns that return 400 errors as they are. This article summarizes what it can do, what it costs, when to choose it, and what to watch for when switching, based on official documentation and announcements that we checked against the originals on October 15, 2026.

Sources: Claude Haiku 5.5 (model page in the official Claude Platform documentation), What's new in Claude Haiku 5.5, Introducing Claude Haiku 5.5 (Anthropic's announcement). Checked October 15, 2026.

1. Basic specs and changes from Haiku 4.5

On the official model list, Haiku 5.5 is one of the current models, alongside Fable 5.1, Opus 5.5, and Sonnet 5.5. Its comparative latency is listed as "Fastest," meaning it is treated as the fastest of the current models. Compared with the earlier Haiku 4.5, the changes are as follows.

Sources: Claude Haiku 5.5, What's new in Claude Haiku 5.5, Thinking, Model deprecations, and Claude Haiku 4.5 (all official Claude Platform documentation, checked October 15, 2026). The knowledge cutoff is the official "reliable knowledge cutoff" value. For a comparison with other companies' models, see List of knowledge cutoff dates for major AI models.

The model ID has no date suffix and no alias. claude-haiku-5-5 is the fixed ID as it is. On Amazon Bedrock, use anthropic.claude-haiku-5-5. On Google Cloud, Microsoft Foundry, and Claude Platform on AWS, use claude-haiku-5-5.

On the official model list, Haiku 4.5 has moved to "Legacy models (still available)," meaning it is no longer updated but can still be used. The status on the Haiku 4.5 page is "Active (legacy)." On the Model deprecations page, which lists planned retirements, the status as of October 15, 2026 is still "Active," and it only states that retirement will not happen before October 15, 2026. No retirement date has been set. Still, it is safest to start anything new on Haiku 5.5 and plan to migrate anything running on Haiku 4.5.

2. Pricing: watch the 100,000-token threshold

Haiku 5.5's pricing is split into two tiers based on the length of the prompt in a single request. Other current models charge the same unit price up to 1M tokens, so this structure is unique to Haiku 5.5.

Source: Pricing (official Claude Platform documentation, checked October 15, 2026). The Batch API is half price for both input and output. For Haiku 5.5, that is $0.05 and $0.25 up to 100,000 tokens, and $0.25 and $1.25 above that.

There is a fixed rule for counting the threshold. According to the official pricing page, prompt length counts all input tokens, including those read from the cache and those written to it. A request that exceeds the threshold is billed at the higher unit price for its entire input, even if part of it hits the cache. The tier is decided per request, and the price of an earlier request never changes later.

Anthropic's announcement says that about 90% of requests to Haiku 4.5 had prompts of 100,000 tokens or less. On average, it says Haiku 5.5 can run about 75% cheaper than Haiku 4.5. The announcement does not explain why the savings are only 75% even though the unit price is one-tenth. However, the token counting and thinking described in the next section are likely contributing factors.

3. What the same job costs: rough estimates for four models

A price table alone makes it hard to see what the bill would look like after switching models, so we calculated one job. The task is sorting 10,000 customer inquiries into categories, assuming 2,000 input tokens and 200 output tokens per item (values counted with Haiku 4.5).

One adjustment is needed here. Haiku 5.5, Sonnet 5.5, and Opus 5.5 count tokens with the new tokenizer introduced with Claude Opus 4.7. According to the official documentation, the same text becomes about 30% more tokens than with Haiku 4.5. For the newer models, we therefore multiplied both input and output by 1.3. The actual increase varies with the content.

Even with 30% more tokens, Haiku 5.5 costs about one-eighth of Haiku 4.5 and one-twentieth of Sonnet 5.5. However, this calculation does not include thinking tokens. Haiku 5.5 has thinking on by default, and thinking is billed as output. For a simple task like sorting, setting effort to low or turning thinking off brings the cost closer to this estimate. Conversely, the more you let it think, the more output you get and the smaller the gap becomes.

To compare pricing across more models, see Pricing and performance comparison of all Claude models. For how caching works and why it sometimes doesn't apply, see our prompt caching article.

4. Performance: numbers from Anthropic's announcement

The following are numbers that Anthropic measured and published itself. They have not been independently verified. The announcement says the details of the testing are in the Haiku 5.5 system card. From the table in the announcement, we pulled out three benchmarks that are easy to compare with Haiku 4.5 and Sonnet 5.5.

There are two points to take from this. First, the gain over Haiku 4.5 is large. On computer use and broad reasoning, the scores are several times those of Haiku 4.5. Second, the gap with Sonnet 5.5 varies greatly by benchmark. The gap is small for computer use, but large for long coding tasks in the terminal.

In the announcement, Anthropic itself says that for complex agentic coding, such as what Terminal-Bench 4.0 measures, Sonnet 5.5 and Opus 5.5 are better suited. It also says Haiku 5.5 is best for narrowly scoped work such as conversation compaction, summarization, and subagents. The announcement also includes comments from companies that tested it early, saying it improved on Haiku 4.5 in correct-answer evaluations and responded faster. These are each company's own evaluations.

5. When to choose Haiku 5.5

Based on the pricing and performance above, here is how to choose. When in doubt, a good rule of thumb is to ask whether the work is narrow and high-volume.

On raising Haiku 5.5's effort too high, the official documentation takes the same view. It recommends using xhigh and max only when your own evaluation confirms a quality improvement. In that case, it also recommends comparing performance, cost, and speed against Sonnet 5.5. For Sonnet 5.5 itself, see our Sonnet 5.5 guide. For Opus 5.5, which sits above it, see our Opus 5.5 guide.

6. Setting thinking and effort

Haiku 5.5 is the first Haiku that lets you choose an effort level (how much it thinks). Its thinking is "adaptive thinking": the model decides whether to think and how much, and effort adjusts how deep that thinking goes. The official documentation recommends the following.

You can also turn thinking off. In the API, sending thinking: {"type": "disabled"} turns it off, but only when effort is high or lower. Combining it with xhigh or max returns a 400 error. According to the official documentation, asking the prompt to "answer right away" did not stop the thinking, so the proper way to reduce thinking is to lower the effort. The documentation also notes that when thinking is off, specifying a JSON output format can cause required tool calls to be skipped.

For the difference between adaptive thinking and the earlier extended thinking, see our adaptive thinking explainer. For what each effort level means, see our article on the effort setting in Claude Code.

7. What to fix when switching from Haiku 4.5

Code that calls Haiku 4.5 through the API may not work just by changing the model ID. From the official migration guide, we list the changes that return 400 errors and the changes that alter results without raising an error.

Sources: Claude Haiku 5.5 migration guide and Prompting Claude Haiku 5.5 (Claude Platform official documentation, checked October 15, 2026).

The thinking configuration can be rewritten as follows (based on the example in the migration guide).

// Before (Haiku 4.5)
{
  "model": "claude-haiku-4-5",
  "max_tokens": 16000,
  "thinking": { "type": "enabled", "budget_tokens": 8000 },
  "messages": [{ "role": "user", "content": "..." }]
}

// After (Haiku 5.5)
{
  "model": "claude-haiku-5-5",
  "max_tokens": 16000,
  "thinking": { "type": "adaptive" },
  "output_config": { "effort": "medium" },
  "messages": [{ "role": "user", "content": "..." }]
}

According to the migration guide, typing /claude-api migrate this project to claude-haiku-5-5 in Claude Code runs the bundled Claude API skill. It swaps the model ID, applies the fixes above in one pass, and outputs a list of items to check by hand. However, you need to verify the token recount and cost estimates yourself. As for the prompts themselves, the official prompting guide says prompts written for Haiku 4.5 generally work well as they are.

8. Using it in Claude Code

In Claude Code, what the alias haiku points to depends on the provider you use and the version of Claude Code. According to the official documentation (Model configuration), as of October 15, 2026, it works as follows.

Even if you use the Anthropic API, an older version of Claude Code will not switch to Haiku 5.5. Update with claude update, then choose it with /model claude-haiku-5-5, or start with claude --model claude-haiku-5-5. If your settings have model set to haiku, resuming a session saved on Haiku 4.5 continues on Haiku 5.5, which is the current target. To keep Haiku 4.5, specify the target for haiku with the environment variable ANTHROPIC_DEFAULT_HAIKU_MODEL.

Using Haiku 5.5 in Claude Code differs from using the API in two ways.

• Thinking cannot be turned off. In the API, thinking can be turned off at effort high or lower. In Claude Code, the Haiku 5.5 thinking toggle displays "Thinking can't be turned off," and MAX_THINKING_TOKENS=0 has no effect either. Adjust how much it thinks with effort.
• The context is 1M tokens from the start. You don't need to add [1m]. By default, conversation compaction runs at about 967K tokens. However, when you use pay-as-you-go API billing, requests with prompts over 100,000 tokens are billed at the higher rate.

Haiku 5.5 is most useful when Claude Code's main model stays Sonnet or Opus, and Haiku handles research, file exploration, and similar tasks as subagents. Anthropic's announcement also says it pairs well as a subagent for Opus 5.5 and Sonnet 5.5. For setup and measured results, see our article on running subagents on a different model.

9. Apps, GitHub Copilot, and cloud

Besides the API and Claude Code, Haiku 5.5 is available in four other places (as of October 15, 2026).

GitHub says that in early testing, Haiku 5.5 matched Claude Sonnet 5 on many coding tasks. This is GitHub's own evaluation. Source: GitHub changelog (October 7, 2026). For Copilot itself, see What is GitHub Copilot?

Summary

1. To choose by price, use Haiku 5.5 for high-volume work under 100,000 tokens, and Sonnet 5.5 or Opus 5.5 for long agentic coding.
2. To switch through the API, handle the items in section 7 first, then recount tokens before estimating costs.
3. To use it in Claude Code, with the Anthropic API on v2.1.293 or later, haiku means Haiku 5.5. Older versions and other providers stay on Haiku 4.5.

FAQ

Q. Can I still use Haiku 4.5?

A. Yes, as of October 15, 2026. The official model list classifies it as "Legacy (still available)." The planned-retirement page only says it will not be retired before October 15, 2026, and no retirement date has been announced yet.

Q. Can Haiku 5.5 replace Sonnet 5.5?

A. It depends on the work. Anthropic itself says Sonnet 5.5 and Opus 5.5 are better suited to complex agentic coding. For narrow, high-volume work, Haiku 5.5 is often enough. The surest way to decide is to try both on your own tasks and compare.

Q. If a prompt exceeds 100,000 tokens, is the whole request more expensive?

A. Yes. If the total input, including tokens read from the cache, exceeds 100,000 tokens, all of that request's input, output, and cache charges are billed at the higher unit price.

Q. Can I save money by turning off thinking?

A. With the API, you can turn it off using thinking: {"type": "disabled"} when effort is high or lower. However, the official documentation recommends lowering the effort first. In Claude Code, Haiku 5.5 thinking cannot be turned off.

Sources

• Claude Platform official documentation: Models overview, Claude Haiku 5.5, What's new in Claude Haiku 5.5, Migration guide
• Claude Platform official documentation: Pricing, Model deprecations, Release notes
• Claude Platform official documentation: Thinking, Effort, Prompting Claude Haiku 5.5
• Claude Code official documentation: Model configuration
• Anthropic: Introducing Claude Haiku 5.5 (October 7, 2026)
• GitHub Changelog: Claude Haiku 5.5 in GitHub Copilot (October 7, 2026)

The official documentation, announcement, and GitHub changelog were all checked against the originals on October 15, 2026. Pricing, availability, and the Claude Code target may change in the future. The benchmarks and company comments come from Anthropic's announcement and were not measured by the author.