Claude Haiku 5.5: price, tiers and uses

What is the Claude Haiku 5.5 price?

The Claude Haiku 5.5 price is $0.10 per million input tokens and $0.50 per million output tokens, for prompts up to 100,000 tokens. A prompt over 100,000 tokens pays $0.50 input and $2.50 output. The model is cheap for short and medium prompts. The long-prompt tier and a new tokenizer both cut into the saving, so measure your own prompts before you plan a budget.

Claude Haiku 5.5 was released on October 7, 2026. Its model ID is claude-haiku-5-5. That is a fixed ID with no date suffix and no separate alias. Prices below are per MTok (one million tokens), taken from Anthropic's pricing page as of October 2026. If the unit itself is new to you, what one million Claude tokens buys turns it into pages of text.

[
  {
    "label": "Haiku 5.5, prompt up to 100K",
    "input_usd": "0.10",
    "output_usd": "0.50",
    "cache_hit_usd": "0.01",
    "cache_write_5m_usd": "0.125",
    "cache_write_1h_usd": "0.20",
    "batch_input_usd": "0.05",
    "batch_output_usd": "0.25"
  },
  {
    "label": "Haiku 5.5, prompt over 100K",
    "input_usd": "0.50",
    "output_usd": "2.50",
    "cache_hit_usd": "0.05",
    "cache_write_5m_usd": "0.625",
    "cache_write_1h_usd": "1.00",
    "batch_input_usd": "0.25",
    "batch_output_usd": "1.25"
  },
  {
    "label": "Haiku 4.5",
    "input_usd": "1.00",
    "output_usd": "5.00",
    "cache_hit_usd": "0.10",
    "cache_write_5m_usd": "1.25",
    "cache_write_1h_usd": "2.00",
    "batch_input_usd": "0.50",
    "batch_output_usd": "2.50"
  },
  {
    "label": "Sonnet 5.5",
    "input_usd": "2.00",
    "output_usd": "10.00",
    "cache_hit_usd": "0.10",
    "cache_write_5m_usd": "2.50",
    "cache_write_1h_usd": "4.00",
    "batch_input_usd": "1.00",
    "batch_output_usd": "5.00"
  }
]

A cache hit (a read from the prompt cache) costs $0.01 per MTok below the line and $0.05 above it. Below the line, a cache write costs $0.125 for a 5-minute cache or $0.20 for a 1-hour cache. The Batch API (asynchronous requests that finish later) takes 50% off input and output. On Haiku 5.5 that means $0.05 input and $0.25 output per MTok for prompts up to 100,000 tokens.

Output costs five times as much as input on every row. Long answers raise the bill faster than long questions, and how input and output tokens are billed on Claude shows how to read that split on your own usage.

Why is Haiku 5.5 priced by prompt length?

Anthropic's pricing page says: "Claude 4.6 and later models (except Claude Haiku 5.5) and Claude Mythos Preview include the full 1M token context window at standard pricing." The same section then says that Haiku 5.5 "is priced by prompt length: a request whose prompt is over 100,000 tokens pays higher prices." Among the 4.6-and-later models, Haiku 5.5 is the exception.

Haiku 5.5 has a 1M-token context window and up to 128K output tokens on the synchronous API. You can send a 400,000-token prompt. You pay the higher row for it. Above the line, input costs five times as much as below it, and so does output. The pricing table picks the row by the length of the prompt, so a long prompt pays the higher rate on all of its input and output.

This matters most for jobs that sit near the line. A prompt that grows slowly, such as a chat history or a retrieval prompt that adds more documents over time, can cross 100,000 tokens without anyone changing code. The cost of that request then jumps.

Do cached tokens count toward the 100,000-token line?

Yes. Anthropic's pricing page answers this under "Long context pricing": "A request's prompt length counts all of its input tokens, including cache reads and cache writes." It adds: "Each request is priced on its own: a request over the threshold pays the higher prices even when part of its prompt is a cache hit, and earlier requests keep the prices they were billed at."

So caching lowers what you pay per token, but it does not move a request below the line. A 120,000-token prompt with 110,000 of those tokens read from the cache still pays the over-100K row. Its cache hits cost $0.05 per MTok, not $0.01, and its output costs $2.50 per MTok. The context-windows page says the same about the window: prompt caching "changes what you pay for those tokens, not whether they count". For a cached conversation that grows past 100,000 tokens, only the requests over the line pay the higher prices. The earlier requests keep the lower prices they were billed at.

None of the worked examples below use caching. To see which side of the line a cached request lands on, read the usage object in each response. It reports input_tokens, cache_read_input_tokens and cache_creation_input_tokens separately, and the three together are the prompt length that sets the price. To decide whether caching pays for itself at all, the break-even point for Claude prompt caching works through the arithmetic.

Is Haiku 5.5 really 90% cheaper than Haiku 4.5?

Per token, yes. Haiku 4.5 lists at $1.00 input and $5.00 output per MTok. Haiku 5.5 lists at $0.10 and $0.50 below the line. That is a 90% cut on the price of one token.

Per job, no. Haiku 5.5 uses the newer tokenizer that Claude 4.7 and later models use. A tokenizer is the part that splits text into tokens, and tokens are what you pay for. Anthropic's notes for Haiku 5.5 say "the same input text produces approximately 30% more tokens on Claude Haiku 5.5 than on Claude Haiku 4.5. The exact increase depends on the content."

More tokens for the same text means a smaller saving than the price table suggests. A token count you measured on Haiku 4.5 is also the wrong input for a Haiku 5.5 estimate. The migration guide says to count your prompts with model set to claude-haiku-5-5 instead of reusing old counts.

What three real jobs cost on Haiku 4.5 and Haiku 5.5

The three jobs below use the same text on both models. Token counts start from Haiku 4.5, and the Haiku 5.5 counts add 30%, the approximate figure from Anthropic's notes. Your content may differ. The examples leave out caching and the Batch API discount. They also leave out thinking tokens. Costs are per 1,000 calls.

[
  {
    "label": "Classification",
    "input_tokens_haiku_4_5": "1,000",
    "input_tokens_haiku_5_5": "1,300",
    "output_tokens_haiku_4_5": "10",
    "output_tokens_haiku_5_5": "13"
  },
  {
    "label": "RAG answer",
    "input_tokens_haiku_4_5": "60,000",
    "input_tokens_haiku_5_5": "78,000",
    "output_tokens_haiku_4_5": "500",
    "output_tokens_haiku_5_5": "650"
  },
  {
    "label": "Long document",
    "input_tokens_haiku_4_5": "150,000",
    "input_tokens_haiku_5_5": "195,000",
    "output_tokens_haiku_4_5": "1,000",
    "output_tokens_haiku_5_5": "1,300"
  }
]

Classification. A short prompt asks for one label, such as a support ticket category. On Haiku 4.5 it counts 1,000 input tokens and 10 output tokens. On Haiku 5.5 the same text counts about 1,300 and 13.

RAG answer. RAG (retrieval-augmented generation) means you paste retrieved documents into the prompt and ask a question about them. This prompt counts 60,000 tokens on Haiku 4.5 and about 78,000 on Haiku 5.5. That is still under the 100,000-token line.

Long document. A long contract or report goes in whole, and you ask for a summary. It counts 150,000 tokens on Haiku 4.5, which fits that model's 200K window. On Haiku 5.5 it counts about 195,000, so it pays the over-100K prices.

[
  {
    "label": "Classification",
    "haiku_4_5_usd": "1.05",
    "haiku_5_5_usd": "0.14"
  },
  {
    "label": "RAG answer",
    "haiku_4_5_usd": "62.50",
    "haiku_5_5_usd": "8.13"
  },
  {
    "label": "Long document",
    "haiku_4_5_usd": "155.00",
    "haiku_5_5_usd": "100.75"
  }
]

A thousand classification calls cost $1.05 on Haiku 4.5 and about $0.14 on Haiku 5.5. A thousand RAG answers drop from $62.50 to about $8.13. A thousand long-document summaries drop only from $155.00 to about $100.75.

[
  {
    "label": "Classification",
    "saving_same_tokens_pct": 90,
    "saving_new_tokenizer_pct": 87
  },
  {
    "label": "RAG answer",
    "saving_same_tokens_pct": 90,
    "saving_new_tokenizer_pct": 87
  },
  {
    "label": "Long document",
    "saving_same_tokens_pct": 50,
    "saving_new_tokenizer_pct": 35
  }
]

If the token counts stayed the same, the first two jobs would save 90%. With 30% more tokens they save about 87%. The long document saves about 35%, because it pays the over-100K prices on Haiku 5.5. Even with the old token count it would save only 50%, because 150,000 tokens is over the line on its own.

There is also a risk between the RAG answer and the long document. A prompt of 80,000 tokens on Haiku 4.5 is about 104,000 tokens on Haiku 5.5 at a 30% increase. That prompt was well under the line on the old model and is over it on the new one. If your prompts sit between about 77,000 and 100,000 tokens on Haiku 4.5, count them again before you switch.

How do I measure my own prompts?

Use the token-counting endpoint. It is free to use, with a requests-per-minute limit that depends on your usage tier. It takes the same body as a normal Messages request and returns the input token count for the model you name. The count is an estimate, and it can differ from the billed count by a small amount.

curl https://api.anthropic.com/v1/messages/count_tokens \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "content-type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-haiku-5-5",
    "system": "You are a support ticket classifier",
    "messages": [{
      "role": "user",
      "content": "My invoice shows the wrong billing address."
    }]
  }'

The response is a small JSON object with one field, input_tokens. Run the same request a second time with the Haiku 4.5 model ID that you call today. The ratio of the two input_tokens values is your real tokenizer increase. Do this with ten or twenty real prompts from your logs, not with one test sentence, because the increase depends on the content.

Then put your own counts into the price table above. Count each prompt on Haiku 5.5 and check which side of 100,000 it lands on. That one check decides which row you pay.

Thinking tokens are billed as output

Haiku 5.5 is the first Haiku model with effort levels. It supports all five: low, medium, high, xhigh and max. The default is medium, both on the Claude API and in Claude Code. Adaptive thinking is on by default, so the model can decide to think before it answers.

Thinking tokens are billed as output tokens, and they count toward max_tokens. Output is the expensive side of every row in the price table. A classification call that thinks for 300 tokens before it writes a 10-token label pays for 310 output tokens. The examples above leave thinking out, so a real bill with thinking on will be higher.

For short, fixed-format jobs you can turn thinking off. That setting works at high effort or below, and the default medium is below it. At xhigh and max the API returns a 400 error instead.

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "content-type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-haiku-5-5",
    "max_tokens": 50,
    "thinking": {"type": "disabled"},
    "messages": [{
      "role": "user",
      "content": "Classify this ticket as billing, bug or other: My invoice shows the wrong billing address."
    }]
  }'

The usage object in the response shows input_tokens and output_tokens. With thinking off, output_tokens should be close to the length of the visible answer. If you want to tune effort instead, the difference between Claude effort levels covers what each step costs you.

What breaks when you swap the model ID?

A cost plan is only useful if the requests still succeed. The Haiku 5.5 migration guide lists changes that return a 400 error on requests that worked on Haiku 4.5:

• Manual thinking with budget_tokens.
• A non-default temperature, top_p or top_k.
• An assistant-message prefill (starting the reply for the model).

The guide names other changes that affect planning. Priority Tier is not supported on Haiku 5.5, so a Priority Tier commitment on Haiku 4.5 does not carry over. Safety classifiers can return stop_reason: "refusal", and there is no server-side fallback, so your code has to handle that stop reason itself. Thinking text is omitted from responses by default. Set thinking.display to "summarized" if you want to see a summary of it.

Where can I use Claude Haiku 5.5?

The model page lists five platforms: the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Bedrock uses the ID anthropic.claude-haiku-5-5. The other four use claude-haiku-5-5. On Anthropic-operated platforms, retirement is not sooner than October 7, 2027. Bedrock and Google Cloud set their own lifecycle dates.

The model docs do not list claude.ai plans for Haiku 5.5. Anthropic does publish a system prompt for Haiku 5.5 on claude.ai and the Claude iOS and Android apps, which describes it as "the fastest model for quick questions". To see whether your plan offers it, open the model picker in the app, or check the model articles on support.claude.com. Plan prices are a separate topic from the API prices on this page.

What is Haiku 5.5 good at?

Anthropic's model overview describes Haiku 5.5 as "for high-volume, latency-sensitive tasks such as classification, extraction, and routing", and lists it as the fastest current model. The launch page names quick, repetitive work such as summaries and compaction (shortening a long conversation so it fits again). It also says Haiku 5.5 "pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work."

The pattern is a narrow task that you run many times. A router that decides which larger model handles a request is a good fit. So is a classifier that tags every support ticket. If those calls run from a script on your own server, controlling AI agent costs on a VPS shows how to put a hard ceiling on what a loop can spend.

When is Sonnet 5.5 the better buy?

Sonnet 5.5 costs $2.00 input and $10.00 output per MTok, at one flat rate across its 1M window. Below 100,000 tokens that is 20 times the Haiku 5.5 price. Above 100,000 tokens it is 4 times. Both models use the newer tokenizer, so the same text counts the same on each, and the comparison is direct.

Choose Sonnet 5.5 when the task is hard enough that Haiku 5.5 fails often. Anthropic's launch page says "Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0", and that Haiku 5.5 "is best suited to more narrowly scoped tasks". A cheap call that you must retry, or check by hand, is not cheap. The effort docs also suggest comparing Haiku 5.5 at xhigh or max against Sonnet 5.5 on cost and speed as well as performance. Thinking tokens are billed as output, so at high effort the real cost per job depends on how much the model thinks. Measure it on your own tasks.

Long prompts also narrow the gap. Above 100,000 tokens Haiku 5.5 is only 4 times cheaper per token. If the job needs careful reading of a long document, Sonnet 5.5 can be the better value for that one job. For a full side-by-side of the model lines, read which Claude model to use: Opus, Sonnet or Haiku. If you are also pricing other vendors, check their own pricing pages for current small-model rates, and see Claude vs ChatGPT API pricing for how to compare the two fairly.

FAQ

How much does Claude Haiku 5.5 cost per million tokens?

For prompts up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens. For prompts over 100,000 tokens it costs $0.50 input and $2.50 output. Cache hits cost $0.01 below the line and $0.05 above it. The Batch API takes 50% off input and output.

Is Claude Haiku 5.5 90% cheaper than Haiku 4.5?

The price of one token is 90% lower below the 100,000-token line. The same text counts approximately 30% more tokens on Haiku 5.5, because it uses a newer tokenizer, so the real saving on a job is smaller. Prompts that land over 100,000 tokens save much less. Count your own prompts with the free token-counting endpoint on both models and compare the two input_tokens values.

What happens when a Haiku 5.5 prompt goes over 100,000 tokens?

The request still runs, because the context window is 1M tokens. It is billed at the higher row: five times the input price and five times the output price of a shorter prompt. Cache reads and cache writes count toward the 100,000 tokens, so a cache hit does not keep a long prompt on the lower prices. Prompts near the line on Haiku 4.5 can cross it on Haiku 5.5, because the newer tokenizer counts more tokens for the same text.

Can I turn off thinking on Claude Haiku 5.5 to save money?

Yes, at high effort or below. Send "thinking": {"type": "disabled"} in the request. The default effort on Haiku 5.5 is medium, so this works without other changes. At xhigh or max effort the API returns a 400 error. Thinking tokens are billed as output tokens, so turning thinking off lowers the cost of short, fixed-format jobs.

Should I use Claude Haiku 5.5 or Sonnet 5.5?

Use Haiku 5.5 for narrow tasks that run many times, such as classification, extraction, routing and subagent work. Use Sonnet 5.5 for complex agentic coding. Also use it for jobs where Haiku 5.5 fails often enough that retries cost more than the price gap. Sonnet 5.5 costs 20 times more per token below 100,000 tokens and 4 times more above it.