Claude Haiku 5.5: Real Cost, Benchmarks and When to Use It

Claude Haiku 5.5 is Anthropic's new small model, released on October 7, 2026. It lists at $0.10 per million input tokens and $0.50 per million output tokens, 90% below Haiku 4.5. That headline is true, but it isn't the whole bill. A new tokenizer, thinking that's on by default and a price step at 100,000 tokens decide what you actually pay.

Key Takeaways

• Claude Haiku 5.5 launched on October 7, 2026, at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens.
• Any prompt over 100,000 tokens pays five times as much on the whole request, not just on the tokens past the line.
• The same text counts as about 30% more tokens than on Haiku 4.5, and thinking is on by default, which is why Anthropic's own estimate of the average saving is about 75% rather than the 90% in the headline.
• Artificial Analysis scores it 43 at max effort, ahead of GPT-6 Luna at 38. But it used about three times as many output tokens to get there.
• Several API habits from Haiku 4.5 now return errors.
• Plan a short migration, not a model-name swap.

What Is Claude Haiku 5.5?

It's the third model in Anthropic's Claude 5.5 family, after Opus 5.5 on September 22 and Sonnet 5.5 on September 28. Haiku is the small, fast tier. Anthropic's model overview lists it as the fastest model in the current lineup.

The jump from Haiku 4.5 is large on paper.

Spec	Claude Haiku 5.5	Haiku 4.5
Context window	1M tokens	200K tokens
Max output	128K tokens	64K tokens
Thinking	Adaptive, on by default	Manual budget, off by default
Effort setting	Low to max, medium default	None
Smallest cacheable prompt	512 tokens	4,096 tokens

It reads text and images and writes text, and its reliable knowledge runs to June 2026, so anything newer than that, such as this month's prices, has to come from your prompt or from a search tool.

It's also the first Haiku model with an effort setting. Five levels, from low to max. Higher effort means more thinking, which usually means better answers, more tokens and more time.

Anthropic pitches it for high-volume work. Think summaries, classification, routing, data extraction, database queries, live customer support and browser use, the kind of jobs that run thousands of times a day, where a second of delay or a fraction of a cent adds up fast. It also suggests running it as a subagent next to Opus 5.5 or Sonnet 5.5 on coding tasks.

Anthropic is just as clear about where it doesn't fit. For complex agentic coding, its announcement points teams to Sonnet 5.5 and Opus 5.5 instead.

How Much Does Claude Haiku 5.5 Cost?

Here's the official price list, per million tokens, from Anthropic's pricing page.

Price per million tokens	Haiku 5.5, prompt up to 100K	Haiku 5.5, prompt over 100K	Haiku 4.5
Input	$0.10	$0.50	$1.00
Output	$0.50	$2.50	$5.00
5-minute cache write	$0.125	$0.625	$1.25
Cache read	$0.01	$0.05	$0.10
Batch API input / output	$0.05 / $0.25	$0.25 / $1.25	$0.50 / $2.50

Under 100,000 tokens, every line is 90% below Haiku 4.5. Over the line, it's 50% below. Anthropic says about 90% of Haiku 4.5 requests stayed under 100,000 tokens, and that mix, plus the tokenizer change below, is why it gives an average saving of around 75%.

Two smaller changes help too. The shortest prompt that can be cached drops from 4,096 tokens to 512, so short system prompts can now be cached. And on the Batch API, Haiku 5.5 can return up to 300,000 output tokens with a beta header.

Anthropic also cut Sonnet 5.5 cache reads in half the same day, from $0.20 to $0.10 per million tokens. It says that lowers the cost of most agentic work on Sonnet 5.5 by about 20%.

Why Isn't the Saving a Full 90%?

Because price per token isn't price per task. Three changes push the token count up, and each one is easy to miss in a quick test.

• First, the tokenizer. Haiku 5.5 uses the same newer tokenizer as Claude 4.7 and later models. Anthropic's what's new page says the same text comes out at about 30% more tokens than on Haiku 4.5, and the exact increase depends on the content.
• Second, thinking is on by default. Adaptive thinking lets the model decide when and how much to think, and those thinking tokens are billed as output. They also count toward your max_tokens limit, so a tight limit can stop a reply after the thinking and before any answer.
• Third, earlier thinking stays in context. In a chat or agent loop, Haiku 5.5 keeps thinking blocks from all earlier turns, and they count as input. Haiku 4.5 kept only the latest turn's blocks, so long conversations grow faster than the tokenizer change alone explains.

None of this cancels the price cut. It just means a fair test measures cost per finished task, on your own prompts, with your own effort setting.

What Is the 100,000-Token Price Line?

It's the detail most likely to surprise a team. Every other Claude 4.6 and later model charges the same rate across the full 1M context window. Claude Haiku 5.5 is the exception.

Anthropic's pricing page puts it plainly. A request whose prompt is over 100,000 tokens "pays higher prices," and the prompt length counts every input token, cache reads and cache writes included. So the higher rate covers the whole request, not just the tokens past the line, even when most of the prompt is a cache hit.

Our math: A prompt of exactly 100,000 tokens costs $0.01 in input. Add one more token. Now the same request costs about $0.05 in input, five times as much, because every one of those 100,001 tokens moves to the higher rate at once. Output on that request also moves from $0.50 to $2.50 per million tokens.

Agent loops feel it most. Each turn resends the conversation, so a long session can cross 100,000 tokens partway through. From then on, every request in that session pays the higher rate on its full prompt and its output, until you trim the context back under the line or start a fresh session.

There are three simple fixes:

• Split long documents into chunks under the line when the task allows it.
• Use thinking block clearing or compaction to keep agent context short.
• Log prompt length per request, so you can see how often you cross.

Long prompts still cost less than on Haiku 4.5. Our math: a document that's 150,000 tokens on the new tokenizer would be about 115,000 tokens on the old one, so it costs about $0.075 in input on Haiku 5.5 against about $0.115 on Haiku 4.5. That's roughly 35% cheaper, but far from 90%.

What Does a Real Workload Cost on Claude Haiku 5.5?

Let's run one example through all of it. This is our own illustrative math, not a measured result, so swap in your own token counts, volumes and effort settings before you use it to make a budget decision.

Say a support team classifies 1 million tickets a month. On Haiku 4.5, each ticket used 2,000 input tokens and 150 output tokens.

Our math, 1M tickets a month	Input tokens	Output tokens	Monthly cost
Haiku 4.5	2.0 billion	150 million	$2,750
Haiku 5.5, thinking off (+30% tokens)	2.6 billion	195 million	$357.50
Haiku 5.5, plus 200 thinking tokens a ticket	2.6 billion	395 million	$457.50

That's 83% to 87% lower, between Anthropic's 75% average and its 90% headline. The gap from 90% comes from the tokenizer and from thinking. If you cache a shared 1,500-token instruction block, input costs fall further, because cache reads cost $0.01 per million tokens.

The picture changes for heavy agent work. A job that runs at max effort, crosses 100,000 tokens and repeats over many turns could save much less. Quality matters too.

How Does Claude Haiku 5.5 Score on Benchmarks?

Strongly for its size on Anthropic's chart, and well ahead of Haiku 4.5.

Vendor claim: These are Anthropic's own figures from its launch page. Sonnet 5.5 is included for scale.

Benchmark	Haiku 5.5	Haiku 4.5	GPT-6 Luna	Sonnet 5.5
GDPval-AA v2.1 (knowledge work, Elo)	1620	735	1437	1840
OSWorld 2.1, offline subset (computer use)	72.4%	15.7%	48.9%	83.9%
Terminal-Bench 4.0 (agentic coding)	39.2%	0.0%	16.4%	70.6%
FrontierCode 1.1 (agentic coding)	46.4%	n/a	42.4%	52.1% (xhigh)
Humanity's Last Exam, with tools	57.4%	18.7%	n/a	64.5%
Chartography, no tools (visual reasoning)	46.4%	6.4%	29.1%	61.6%

The big gap to Sonnet 5.5 on Terminal-Bench 4.0, 39.2% against 70.6%, backs up Anthropic's own advice. Hard coding work still belongs on a larger model.

Independent data: Artificial Analysis gives a more mixed view in its launch write-up. Its Intelligence Index puts Haiku 5.5 at 43 at max effort and 38 at high effort. That's ahead of Gemini 3.8 Flash at 41 and GPT-6 Luna at 38, just behind Kimi K3 at 44, and well below Sonnet 5.5 at 56.

On its own Terminal-Bench 4.0 run, Artificial Analysis measured 33%, against 13% for GPT-6 Luna and 0% for Haiku 4.5. Lower than claimed. But the direction holds.

Where Does It Fall Short?

Three results stand out from the independent tests, and each one is worth checking against your own workload before you route real traffic to the model. They don't sink the model, but they tell you what to check.

It thinks a lot at max effort. Artificial Analysis says Haiku 5.5 used about 162,000 output tokens per Index task at max effort, roughly three times GPT-6 Luna's 50,000. At high effort it used about 55,000. Our math: at the same $0.50 rate, that's about $0.08 of output per task at max against about $0.025 for Luna, before input costs.

So the fair match-up is Haiku 5.5 at high effort against Luna at max. Both score 38, with similar token counts. Max effort buys 5 more points at about three times the output.

It knows less, but guesses less. On AA-Omniscience, Haiku 5.5 got 36% of answers right, below Gemini 3.8 Flash at 55% and Luna at 44%. But its hallucination rate was 40%, lower than Gemini 3.8 Flash at 55% and Luna at 77%. For lookups, pair it with search or your own data rather than its memory.

It refused too often in early testing. Haiku 5.5 scored 35% on AutomationBench-AA, against 53% to 60% for Luna, Gemini 3.8 Flash and GLM-5.3 Flash. Artificial Analysis says a pre-release safety bug caused over-refusal, and it plans to re-run the test after Anthropic's fix.

Hands-on reports point the same way. A Decrypt reviewer found it very fast but got a wrong answer to a simple logic question. Fast and cheap isn't the same as checked.

How Does Claude Haiku 5.5 Compare With GPT-6 Luna?

At list price, they're tied. Both cost $0.10 and $0.50 per million tokens, based on Anthropic's page and OpenRouter's listing for Luna.

Factor	Claude Haiku 5.5	GPT-6 Luna
List price per million tokens	$0.10 / $0.50 under 100K	$0.10 / $0.50
Intelligence Index (Artificial Analysis)	43 max, 38 high	38 max
Output tokens per Index task at max	About 162,000	About 50,000
Terminal-Bench 4.0 (Artificial Analysis)	33%	13%
AA-Omniscience hallucination rate	40%	77%

Haiku 5.5 leads on agent and computer-use tests, and it guesses less often. Luna is leaner with tokens at max effort and did better on AutomationBench-AA during the refusal bug. If you already run Luna in production, put both models through the same set of real tasks, at matched effort, before you move any traffic from one to the other. Our Sonnet 5.5 vs GPT-6 Luna comparison covers the larger tier.

What Breaks When You Migrate?

More than a model name, because Anthropic's migration guide lists several habits that worked on Haiku 4.5 and now return errors on Haiku 5.5.

• Manual thinking budgets: setting budget_tokens returns an error. Use adaptive thinking and the effort setting instead.
• Sampling settings: temperature, top_p and top_k can return an error. Anthropic says to leave them out.
• Assistant prefill: a conversation that ends with an assistant turn returns an error. End with a user turn and ask for the format in the prompt.
• Old computer-use tool: computer_20250124 returns an error. Move to the newer tool version for your platform.
• Structured outputs on Amazon Bedrock: these aren't available there. Describe the format in the prompt and check the output in code.

Some changes won't throw errors but can still break things:

• Responses can begin with thinking blocks, so read content by its type, not by position.
• Thinking text is hidden by default unless you ask for a summary.
• Thinking blocks only work in the account that produced them, which matters if you replay stored chats.

Safety classifiers can also decline a request with a refusal stop reason. On Haiku 5.5, server-side fallback isn't available, so your own code has to catch it and route the request elsewhere. Security teams, take note. Anthropic says its cyber rules allow more defensive work than Sonnet 5.5 does, but penetration testing stays blocked.

Where Does Claude Haiku 5.5 Fit in Your Stack?

As the high-volume worker under a larger model, in most setups we see. The customer reports Anthropic published point the same way. They're vendor-chosen quotes, though.

• Asana said task completion latency fell more than 30% compared with its current model.
• HubSpot said Haiku 5.5 scored 92.8% on its internal suite, averaged over three runs, the best it had seen.
• AlphaSense reported 0.84 against 0.76 for Haiku 4.5 across 400 queries.
• Box said it scored 11 points higher than Haiku 4.5 at about half the latency.

We'd sort work into three lanes, with a larger model planning the work and Haiku doing the many small reads and checks underneath it.

Lane	Model and effort	Example tasks
High volume	Haiku 5.5, low or medium	Classification, routing, tagging, extraction, short summaries
Subagent steps	Haiku 5.5, high	Reading filings, checking pages, browser tasks
Hard work	Sonnet 5.5 or Opus 5.5	Planning, hard coding, anything where one wrong answer is costly

Move a task down a lane only when your tests say it holds up. Our Sonnet 5.5 guide and Opus 5.5 benchmark breakdown cover those tiers.

Whatever you choose, keep the model behind a thin routing layer. Prices and rankings shift every few weeks, and our model portability lesson explains why that flexibility pays off.

How Should Teams Test Claude Haiku 5.5?

Run a one-week trial on real work before you switch traffic. Here's the plan we use.

1. Pull 200 to 500 real tasks from your logs, with known good answers, and recount their tokens with the token counting endpoint.
2. Try low, medium and high effort, and record accuracy, latency and cost per task.
3. Log prompt length and refusals on every request, and flag any that cross 100,000 tokens.
4. Compare against your current model on the same tasks, and switch only the lanes that pass.

Start with the lane that has the most volume and the least risk. That's usually classification or routing, where errors are easy to catch and the savings show up fast.

Should You Switch From Claude Haiku 4.5?

For most high-volume work, yes, after a short test. The price cut is real. So is the quality jump, which shows up both on Anthropic's own chart and on the independent tests, and Anthropic also says the model shows far fewer signs of misaligned behavior than Haiku 4.5.

• Switch first: short classification, routing and extraction jobs that stay well under 100,000 tokens.
• Check before switching: long-document jobs and agent loops, where the price line and growing context can eat the saving.
• Check twice: anything that relies on prefill, fixed temperature or security tooling.

Haiku 5.5 will stay available until at least October 7, 2027, so there's no rush. Take the week to test it properly.

How Van Data Team Helps Teams Route Work Across Claude Models

We help teams decide which model handles which task, based on their own data. That means eval sets built from real tasks, cost-per-task tracking, prompt-length alerts and routing rules with clear quality gates.

Claude Haiku 5.5 is a fast, strong small model at a very low list price. Whether it saves you 87% or 35% depends on how your prompts look. If you want help measuring that, our AI governance guide is a good place to start.