Claude Haiku 5.5 Costs $0.10 per Million Tokens: Which Jobs to Move First

Claude Haiku 5.5 launched on October 7, 2026, at $0.10 per million input tokens and $0.50 per million output tokens. That is 90% below Claude Haiku 4.5, for prompts up to 100,000 tokens.

Anthropic shipped three related changes the same day. Sonnet 5.5 cache reads dropped by half. Max and Team plans now include monthly API credits, nonprofit Team plans included. Claude Code made Haiku 5.5 its default Haiku model. This post covers the specs, the price limits, how Haiku 5.5 compares with OpenAI and Google, and which high-volume jobs agencies and nonprofits can move first.

The short version: Haiku 5.5 matches GPT-6 Luna’s list price token for token, has a 1M-token context window, and still accepts forced tool use. The low price stops at 100,000 tokens per prompt. Haiku 5.5 also thinks by default, uses about 30% more tokens for the same text, and rejects temperature. Start with classification and tagging, turn thinking off where you can, and measure cost on real traffic. All figures are as of October 10, 2026.

What Anthropic shipped with Claude Haiku 5.5

Anthropic’s Haiku 5.5 announcement calls it the cheapest and fastest small model the company has released. The model ID is claude-haiku-5-5, a fixed ID with no date suffix and no separate alias.

It has a 1M-token context window and 128k max output tokens. It takes text and images and returns text. Its reliable knowledge cutoff is June 2026. It is the first Haiku model with an effort setting, and its default effort is medium.

Prices come from Anthropic’s pricing page. Haiku 5.5 runs on the Claude API, Amazon Bedrock (anthropic.claude-haiku-5-5), Google Cloud, Microsoft Foundry and Claude Platform on AWS.

Anthropic also published benchmark scores. On its offline subset of OSWorld 2.1, a computer-use test, it reports 72.4% for Haiku 5.5, 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna. These are Anthropic’s own figures. Anthropic also says Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding.

The $0.10 price stops at 100,000 tokens

The headline price has a ceiling. When a prompt runs over 100,000 input tokens, the whole request moves to a higher tier: $0.50 input, $2.50 output and $0.05 for cache reads.

Two details matter here. First, cache reads and cache writes count toward the 100,000. A chatbot with a long cached system prompt and a long conversation can cross the line without anyone noticing. Second, each request is priced on its own. One long request does not change the price of the next short one.

According to the pricing page, Haiku 5.5 is the only Claude model from the 4.6 generation on without standard pricing across its full 1M window. Above 100,000 tokens it still costs half of what Haiku 4.5 costs. It just no longer costs a tenth.

The tokenizer changes the math too. Anthropic’s migration guide says the same text produces about 30% more tokens on Haiku 5.5 than on Haiku 4.5. A 90% per-token cut on 30% more tokens works out to about 87% less for the same text, by our arithmetic. Anthropic’s own estimate is that Haiku 5.5 costs around 75% less to run on average. Thinking tokens are the likely gap, and you can control those.

How Claude Haiku 5.5 compares with GPT-6 Luna and Gemini

We checked the three vendors’ own pricing pages on October 10. Each vendor counts tokens differently, so treat per-token comparisons as rough.

On short prompts, Haiku 5.5 and Luna have identical list prices, including the $0.125 cache write. OpenAI’s GPT-6 Luna model page lists a 1,050,000-token context window and 128,000 max output. We covered Luna’s launch in our GPT-6 Sol and Luna post.

The difference is long prompts. Between 100,000 and 272,000 tokens, Luna stays at $0.10 while Haiku 5.5 costs $0.50. Above 272,000, Luna costs $0.20 input and Haiku 5.5 still costs $0.50.

Google’s Gemini pricing page lists no prompt-size tier for these two models. Both have a 1,048,576-token input window and 65,536 max output. Gemini 3.8 Flash costs more per token, and our Gemini 3.8 Flash cost analysis found it costs more per task than 3.7 Flash at the same price. All three vendors offer batch processing at 50% off.

Five jobs to move to Claude Haiku 5.5 first

Anthropic’s model overview describes Haiku 5.5 as built for high-volume classification, extraction and routing. That lines up with the work agencies run quietly for nonprofit and association clients. These are the five we would test first.

Classification. Sorting support emails, membership inquiries or donation notes into a fixed set of labels. The prompts are short, the outputs are tiny, and a wrong label is easy to spot in a sample.

Routing. Deciding which queue, staff member or model should handle a request. Anthropic’s announcement pairs Haiku 5.5 with Sonnet 5.5 or Opus 5.5 as a coding subagent. The same split works for routing: a cheap first pass sends only the hard cases to the larger model.

Content tagging. Applying taxonomy terms to articles, resources or archive items in a CMS. This is a good fit for the Batch API at $0.05 input and $0.25 output, run overnight.

Chat triage. Answering simple questions and handing off the rest. Watch the 100,000-token line here, since long histories and cached prompts count toward it.

Form processing. Pulling fields out of applications, event registrations or grant forms. Haiku 5.5 still accepts forced tool_choice, unlike Sonnet 5.5, so a tool with fixed fields keeps working. The response starts with the tool call and has no thinking block.

Keep long-document work, such as summarizing a full annual report, on a test track. Those prompts land in the higher tier, where Luna and Gemini Flash-Lite cost less per input token.

What breaks when you move from Haiku 4.5

Haiku 4.5 code can break on Haiku 5.5. The migration guide lists these 400 errors.

• Sampling parameters. Any `top_k` value errors. `temperature` must be 1, and sending both `temperature` and `top_p` errors. Many classification scripts set temperature to 0. Remove it.
• Thinking budgets. `budget_tokens` is rejected. Use adaptive thinking with an effort level.
• Prefill. A conversation must end with a user message. Use structured outputs or a tool with enum fields for fixed formats.
• Old computer use tool. `computer_20250124` is rejected on the Claude API and Google Cloud.
• Edited history. Changing the system prompt, tools or earlier turns while sending thinking blocks back can return an error.

Two gaps are platform-specific. Structured outputs are not available for Haiku 5.5 on Amazon Bedrock, so validate output in your own code there. Priority Tier is not supported on Haiku 5.5.

Thinking is the change most likely to affect your bill. Haiku 4.5 did not think unless asked. Haiku 5.5 thinks by default at medium effort, and thinking tokens bill as output. Anthropic’s effort guide says Haiku 5.5 accepts thinking: {"type": "disabled"} at high effort or below. It recommends low effort for simple, high-volume requests.

Haiku 4.5 is still active, for now

Anthropic’s deprecations page lists Haiku 4.5 as active, with retirement "Not sooner than October 15, 2026." No deprecation has been announced as of October 10.

The page sets no firm date. It only says retirement will not come before October 15. For comparison, Sonnet 4.5’s deprecation came on September 30 with a November 30 retirement, two months of notice. We covered that in our Claude Sonnet 5.5 post.

Many agencies have Haiku 4.5 inside plugins, chat widgets and tagging scripts that nobody has opened in a year. Finding them now costs less than finding them after a deprecation notice. OpenAI teams have a parallel deadline: GPT-5.4 Nano ends April 1, 2027, as covered in our OpenAI API shutoff post.

The other October 7 changes

Sonnet 5.5 cache reads halved. Anthropic’s release notes say Sonnet 5.5 cache reads dropped from $0.20 to $0.10 per million tokens. Cache writes and other prices did not change. Anthropic estimates this cuts Sonnet 5.5’s cost on most agentic tasks by around 20%.

Monthly API credits for Max and Team. Max 5x gets $100 a month and Max 20x gets $200. Team gets $20 per Standard seat and $100 per Premium seat, pooled and capped at $500. Anthropic’s API credits page says discounted Team plans, including Nonprofit, qualify.

You claim credits on claude.ai, under billing settings, and they appear in the Console as promotional credits. They expire at the end of each billing cycle. They work on the Claude API but not on Bedrock, Google Cloud, Foundry or in Claude Code. At the short-prompt rate, $100 covers about a billion Haiku 5.5 input tokens.

Claude Code default. The Claude Code changelog for version 2.1.293 lists Haiku 5.5 as the default Haiku model on the Anthropic API. For where Haiku fits next to the larger models, see our Claude Opus 5.5 post and our tier selection guide.

What to test first

1. Pick one job. Classification or tagging is the easiest to grade.
2. Build a labeled sample of 200 to 500 real items with known correct answers.
3. Change the model ID to `claude-haiku-5-5` and remove `temperature`, `top_p`, `top_k` and prefill.
4. Run two settings: thinking disabled at `low` effort, and the default `medium`. Compare accuracy and output tokens.
5. Count tokens with the new model ID, and log any request over 100,000.
6. Run the same sample on Haiku 4.5 and compare cost per correct answer, not cost per token.
7. Move to Batch for any job that can wait overnight.

Frequently Asked Questions