Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Pricing

Claude Haiku 5.5 Costs $0.10 per Million Input Tokens. Past 100,000 in One Prompt, the Whole Request Costs Five Times More.

October 19, 2026 9 min read

On Wednesday, October 7, Anthropic released Claude Haiku 5.5, the new version of its small, fast model. The headline is the price. Haiku 4.5 cost $1 per million input tokens and $5 per million output tokens. Haiku 5.5 costs $0.10 and $0.50. Anthropic says it runs around 75 percent cheaper on average, and pitches it for high-volume work: summaries, compaction, classification, and subagents working for Opus 5.5 or Sonnet 5.5.

Look at the price table on the launch page and every Haiku 5.5 cell holds two numbers. Input reads "$0.10 / $0.50". The second number is the price for prompts over 100,000 tokens. Simon Willison put it plainly the same day: beyond 100,000 tokens, "the price increases 5x".

The pricing documentation spells out how. "Claude Haiku 5.5 is priced by prompt length: a request whose prompt is over 100,000 tokens pays higher prices." The same page says Claude 4.6 and later models get their full 1M-token context window at standard pricing, "except Claude Haiku 5.5". Haiku 5.5 has a 1M-token window too. The cheap part of it is the first tenth.

I build a LegalTech product on Claude, and I built BidForge, an AI that reads French public tender files for small businesses. Long documents are the normal case in both. So I read the pricing page and the migration guide before the benchmarks. Here is what they say, with the arithmetic.

A cliff, not a tier

In a tax bracket, only the part above the threshold pays the higher rate. This is not that. Once the prompt is over 100,000 tokens, every token in the request pays the higher price, output included.

Above the line, input goes from $0.10 to $0.50 per million tokens, output from $0.50 to $2.50, five-minute cache writes from $0.125 to $0.625, cache reads from $0.01 to $0.05. The Batch API halves both columns ($0.05 and $0.25 under the line, $0.25 and $1.25 over it) but keeps the line. US-only inference adds its 1.1x multiplier on top, the docs say, including on the higher prices.

Take a request with a 99,000-token prompt and 2,000 tokens of output. Input: 99,000 × $0.10 / 1M = $0.0099. Output: 2,000 × $0.50 / 1M = $0.001. Total: $0.0109.

Add 2,000 tokens to the prompt. Input: 101,000 × $0.50 / 1M = $0.0505. Output: 2,000 × $2.50 / 1M = $0.005. Total: $0.0555.

Two percent more prompt, about five times the bill. Over 100,000 such requests, $1,090 becomes $5,550.

The output side deserves a second look. Thinking is on by default on Haiku 5.5, with effort at medium, and the docs say thinking tokens are billed as output tokens even when the thinking text is not returned to you. Above the line, those tokens cost $2.50 per million too.

The cache does not protect you

On most Claude workloads, the prompt cache is the first cost lever. Here it does not move the line. The docs: "A request's prompt length counts all of its input tokens, including cache reads and cache writes." And "a request over the threshold pays the higher prices even when part of its prompt is a cache hit."

Those two sentences came after the launch. The Internet Archive's copies of the pricing page from October 7 and 8 describe the threshold in one sentence: "a prompt of over 100,000 tokens pays higher prices." The copy from October 9 has the cache rule. If you read the page on launch day and assumed cached tokens did not count, the page has since said otherwise.

So the same 90,000 cached tokens cost 90,000 × $0.01 / 1M = $0.0009 to read in a request under the line, and 90,000 × $0.05 / 1M = $0.0045 in a request over it. Caching still cuts the input price by 90 percent on either side. It just cannot keep you on the cheap side.

Three ways a prompt crosses the line without anyone touching it

The tokenizer. Anthropic's launch post says Haiku 5.5 "uses slightly more tokens per task." The migration guide is more specific: "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." Simon Willison measured about 1.25x on one long prompt and called it "a hidden price increase." The consequence: 100,000 / 1.3 is about 77,000. A prompt that measured 77,000 tokens on Haiku 4.5 can land right at the line on Haiku 5.5. The guide tells you to recount with the new model and redo cost estimates "including its higher prices for long prompts."

Thinking that stays in context. On Haiku 5.5, "Thinking blocks from earlier assistant turns stay in context and count as input tokens, where Claude Haiku 4.5 kept only the latest turn's." A multi-turn agent therefore grows faster than the tokenizer change alone explains. Thinking block clearing exists to trim them.

Images and PDFs. Haiku 5.5 uses the high-resolution image tier. The guide's example: a 2,000 by 1,500 pixel image costs about 2.5 times as many visual tokens as on Haiku 4.5. PDFs are processed as text plus an image of each page, and Anthropic's PDF documentation estimates 1,500 to 3,000 text tokens per page. On text alone, that puts the line somewhere between about 33 and 67 pages, before the page images are counted.

The default that sits on the line

Anthropic pitches Haiku 5.5 for compaction, the job of summarizing a long conversation so it fits again. The SDK's client-side compaction helper has a default threshold of 100,000 tokens, counted as input plus cache writes plus cache reads plus output. When the threshold is exceeded, it injects a summary request into the conversation, and the summary model defaults to the main model.

If that model is Haiku 5.5, the summary request carries the whole history that just went past 100,000 tokens. The one call meant to save money is priced above the line. The default trigger for tool result clearing in context editing is also 100,000 input tokens. On Haiku 5.5, I would move both a margin below the price line.

How it compares

Under the line, Simon Willison notes, Haiku 5.5 exactly matches the price of OpenAI's GPT-6 Luna: $0.10 and $0.50. Luna has a threshold too, at 272,000 input tokens, where OpenAI's model page says prompts are "priced at 2x input and cache rates and 1.5x output for the full request": $0.20 and $0.75. Between 100,000 and 272,000 tokens, Luna's list price is a fifth of Haiku's. Above that, Haiku costs 2.5 times more on input and about 3.3 times more on output. Each vendor counts tokens with its own tokenizer, so the only fair comparison is your own documents, counted by both.

Inside Anthropic's range, the long-prompt case is still a good deal. Above the line, Haiku 5.5 costs a quarter of Sonnet 5.5's $2 and $10, which stay flat across the full window. And Anthropic's footnote prices Haiku 5.5 90 percent below Haiku 4.5 for requests up to 100,000 tokens and 50 percent below it for requests over. It adds that 90 percent of Haiku 4.5 requests fell in the first group, and that its average of around 75 percent less already accounts for the new tokenizer. That is Anthropic's traffic. Your number depends on where your own requests fall.

What I would do before switching

Measure the distribution, not the average. Log the prompt length of every request (uncached input, cache writes and cache reads together) and see how much of your Haiku 4.5 traffic sits between 77,000 and 100,000 tokens, and how much is already over. The token counting endpoint, with the model set to claude-haiku-5-5, gives you the new count before you send.

Put a guard in the client. When a counted prompt is about to cross the line, decide in code: trim it, split it, or route it to another model. Not by accident, on the invoice.

Split when the task allows it. Take a 150,000-token document and a 3,000-token answer. In one call: 150,000 × $0.50 / 1M + 3,000 × $2.50 / 1M = $0.0825. In two calls of 75,000 tokens with 1,500 tokens of output each, plus a merge call of 3,000 tokens in and 1,500 out: about $0.0176. Roughly a fifth. That works for extraction and per-section summaries. It does not work for a question that needs the whole document in view at once, and that is the case where I would pay the higher price, or pick another model, on purpose.

For a pipeline like BidForge's, the first question I would ask of each step is whether it needs the whole tender file in view, or only the one document it is reading.

Set compaction and clearing thresholds a margin below 100,000, around 80,000, so the summary call itself is billed under the line.

And compare cost per task, not price per token. When Opus 5.5 and GPT-6 Sol launched 90 minutes apart on September 22, I wrote that cost per task had become the number to read. Haiku 5.5 adds a rule: on this model, the length of the prompt is part of the price.

Haiku 5.5 is a tenth of the price on the requests that fit. The engineering work is making sure they fit.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk