Disclosure: this guide is based on Anthropic’s announcement, pricing and migration documentation, and third-party benchmark reporting. We have not run Claude Haiku 5.5 in a long production deployment, and figures are attributed to their sources.
Claude Haiku 5.5 launched on October 7, 2026, at $0.10 per million input tokens and $0.50 per million output tokens. Haiku 4.5 lists at $1 and $5. That is a 90% cut per token, and Anthropic calls it the cheapest, fastest and most capable small model it has ever released.
The honest answer is that it really is much cheaper, but not quite as much cheaper as the headline. A new tokenizer turns the same text into roughly 30% more tokens, and prompts over 100,000 tokens pay five times the base rate on the entire request. Budget for both and Claude Haiku 5.5 is a superb workhorse for high-volume tasks. Ignore them and your forecast will be wrong by a third or more.
Table of Contents
- What Claude Haiku 5.5 Is For
- Claude Haiku 5.5 Pricing and the 100K Cliff
- The Tokenizer Footnote
- Effort Levels Arrive on Haiku
- Claude Haiku 5.5 vs GPT-6 Luna
- Jobs Claude Haiku 5.5 Does Well
- Five Costly Budgeting Mistakes
- Designing a Pipeline Around a Small Model
- Migrating to Claude Haiku 5.5
- Who Should Use What
- FAQ
- Verdict
- Sources and Further Reading

What Claude Haiku 5.5 Is For
Claude Haiku 5.5 is the third model in Anthropic’s 5.5 family, after Opus 5.5 on September 22 and Sonnet 5.5 on September 28. It is the small one. Anthropic positions it for quick, repetitive work such as summaries, context compaction, database queries and classification, and for sub-agents that sit underneath a bigger model doing the heavy thinking.
It is also Anthropic’s fastest model, which matters for real-time customer support and browser automation, where every second of latency is felt by a person waiting. The model reads text and images and writes text.
Availability is broad from day one: the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. The ID is claude-haiku-5-5. Anthropic commits to keeping it available until at least October 7, 2027, so it is safe to build on for a year.
Claude Code switched its default Haiku model to Claude Haiku 5.5 on the Anthropic API, according to release notes cited by Penchan. That matters because Claude Code quietly uses the small model for background chores, so many developers are already running it without having chosen it.
Claude Haiku 5.5 Pricing and the 100K Cliff
Haiku 5.5 pricing is tiered by prompt length, and it is the only current Claude model priced that way. Here is the full table.
| Prompt length | Input per 1M | Output per 1M |
|---|---|---|
| Up to 100,000 tokens | $0.10 | $0.50 |
| Over 100,000 tokens | $0.50 | $2.50 |
| Haiku 4.5, for comparison | $1.00 | $5.00 |
The cliff is the detail to plan around. Cross 100,000 prompt tokens and the whole request is billed at the higher rate, not just the portion past the line. A 99,000-token prompt and a 101,000-token prompt differ by 2% in size and roughly 400% in input cost.
Anthropic’s own justification is that requests under 100,000 tokens made up around 90% of traffic to the previous Haiku. So for most users the low tier is the only one that matters. Even above the cliff, Claude Haiku 5.5 is still half the price of Haiku 4.5 per token, so long prompts are not a disaster. They are just not the bargain the headline suggests.
The averaging explains why different outlets reported different discounts. VKTR and others reported 90%, comparing the low tier to Haiku 4.5. AI Weekly reported roughly 75% on average, which folds in the higher tier and the token count. Both are correct; they measure different things.
The Tokenizer Footnote
The tech site Mixed found the detail that the launch post did not lead with. Anthropic’s what’s-new page, its migration guide and its pricing page all state that the same text produces approximately 30% more tokens on Claude Haiku 5.5 than on Haiku 4.5. Nothing was hidden. It just was not in the headline.
Here is what it does to the math. If a job used one million tokens on Haiku 4.5, it will use about 1.3 million on Claude Haiku 5.5. At $0.10, that is $0.13 instead of $1.00 on input. The real saving is about 87%, not 90%. Still enormous. Just not the number on the slide.
The bigger risk is not the price but the limits. If you have hard caps on tokens per request, chunk sizes tuned to a token budget, or context-length checks in your code, all of those shift by 30%. A chunk that fitted comfortably before might now spill past a limit. Recount before you migrate.
Effort Levels Arrive on Haiku
Claude Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, the dial Anthropic now uses in place of a manual thinking budget. That is a bigger change than it sounds. Previous small models were essentially one speed; now you can trade tokens for capability per request.
For classification and extraction, low effort is usually right. For anything involving a few steps of reasoning, such as deciding which tool an agent should call next, medium is a sensible start. High and max exist, but if you find yourself needing them often, Sonnet 5.5 at medium may be cheaper per successful task.
The effort setting also shapes latency. Artificial Analysis data cited by Codersera shows Haiku at high effort producing its first chunk in about 26 seconds on its benchmark suite, which is fine for a batch job and far too slow for a chat window. Interactive features want low or medium.
Claude Haiku 5.5 vs GPT-6 Luna
OpenAI’s GPT-6 Luna, released September 22, is the obvious rival, and on base price the two match exactly at $0.10 and $0.50. The differences sit in the details.
| Dimension | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Base price per 1M | $0.10 / $0.50 | $0.10 / $0.50 |
| Price cliff | Above 100K tokens: $0.50 / $2.50 | Above 272K tokens: $0.20 / $0.75 |
| AA Index at max effort | 43 at $0.21 per task | 38 at $0.07 per task |
| AA Index at high effort | 38 at $0.08 per task | 33 at $0.03 per task |
| Longest cache option | 1 hour | 12 hours |
The figures come from Codersera’s compilation of Artificial Analysis Intelligence Index data from late October. The pattern is clear. Claude Haiku 5.5 is smarter per effort level, and Luna is cheaper per task. At high effort, Haiku matches Luna’s max-effort score at about the same cost while responding much faster.
The cliff placement favours Luna for long documents. Prompts between 100,000 and 272,000 tokens are five times the base rate on Haiku and still the base rate on Luna. If your workload lives in that band, Luna wins on cost almost by default.
The cache window favours Luna too, for workloads that reuse a long prompt across a whole working day. For bursty traffic within an hour, the difference disappears.
Jobs Claude Haiku 5.5 Does Well
Routing and classification
Deciding which department a ticket belongs to, whether an email is spam, or which tool an agent should call. These are short prompts with short outputs, exactly where the low tier shines.
Sub-agents under a larger model
Anthropic explicitly pitches Claude Haiku 5.5 as a sub-agent for Opus 5.5. A planner model decides what to do, and Haiku carries out the cheap parts: reading a file and summarising it, checking a log for errors, reformatting output. The planner never sees the raw bulk.
Context compaction
Long agent sessions eventually need their history summarised to stay within limits. Handing that job to Haiku rather than the main model saves money on every compaction, and summaries are a task small models handle well.
Browser and computer use
The New Stack noted big gains on computer-use and command-line benchmarks. Speed matters here, because browser agents take many small actions and each one waits on the model.
Structured extraction
Pulling fields from invoices, forms or support transcripts into JSON. Keep documents under the cliff and this is one of the best value uses of the model.
Five Costly Budgeting Mistakes
- Using the old token counts. Multiply your Haiku 4.5 token volumes by about 1.3 before applying new prices.
- Letting prompts drift past 100K. Add a check that warns or splits requests before they cross the line. A few accidental long prompts can erase a month of savings.
- Defaulting to high effort. Most Haiku tasks need low or medium. Higher effort spends output tokens, the expensive half of the bill.
- Ignoring cache structure. Put stable instructions first so repeated calls hit the cache. The one-hour window means bursty jobs benefit most.
- Using Haiku for hard reasoning. If you need max effort and retries to get a right answer, a mid-tier model at medium may cost less per success.
There is also a bonus for subscribers. With the Claude Haiku 5.5 launch, Anthropic began rolling out monthly Claude Platform API credits for Max and Team plans: $100 for Max 5x, $200 for Max 20x, $20 for a standard Team seat and $100 for an advanced seat, with a $500 monthly cap per team. At Haiku prices, $100 buys a lot of tokens.
Designing a Pipeline Around a Small Model
The biggest savings rarely come from swapping one model for another. They come from redesigning the pipeline so that the expensive model handles only the decisions that need it. A cheap, fast model makes that redesign worthwhile.
A common pattern looks like this. Incoming requests hit a small model first, which classifies them and handles the easy majority directly: password resets, order status, simple factual questions. Only the requests it flags as complex get passed to a larger model. If the small model is right about difficulty most of the time, the average cost per request drops sharply, and the larger model spends its time on work that justifies its price.
The trick is measuring the escalation rate honestly. Track how often the small model escalates, how often it should have escalated but did not, and how often the larger model was called for something trivial. Those three numbers tell you whether the routing logic is working. Without them, you are guessing.
Another pattern is pre-processing. Before a long document reaches an expensive model, a small one can strip boilerplate, extract the relevant sections and drop everything else. Cutting a 200-page contract down to the 15 pages that matter saves far more than any per-token discount, and it also keeps the request under length-based price cliffs.
Finally, keep the small model’s prompts short and specific. Small models degrade faster than large ones when instructions are vague or overloaded. One task per call, clear output format, and a worked example usually beat a long list of rules.
Migrating From Haiku 4.5 to Claude Haiku 5.5
For most teams, moving to Claude Haiku 5.5 is a model-name change plus an afternoon of checks. Anthropic says Managed Agents users need only the name change, and the migration guide covers the rest. Here is the order I would do it in.
First, re-run your token counting. Any code that estimates cost or enforces limits by counting tokens needs the new tokenizer, or every estimate will be about 30% low. This is the step people skip, and it is the one that causes surprise invoices.
Second, re-check chunk sizes for retrieval pipelines. If your document chunks were sized to a token budget, they now hold less text per chunk under the same limit, or overflow if the limit is fixed. Either adjust the budget or re-chunk.
Third, set an explicit effort level. Leaving it on the default is fine for a first pass, but measure. Many classification jobs will run perfectly well on low, and that is where Claude Haiku 5.5 is at its cheapest and fastest.
Fourth, watch the first week of bills by prompt size. Plot how many requests land above 100,000 tokens. If the answer is more than a handful, find out why. Usually it is one job that concatenates too much context, and fixing that single job recovers most of the expected savings.
Finally, keep Haiku 4.5 reachable for a short period. Rolling back should take one configuration change. The newer model is better on nearly every benchmark, but your workload is the only benchmark that pays the bills, and a quick comparison on real traffic settles the question for good.
Who Should Use What
| If you are | Start with | Why | Link |
|---|---|---|---|
| Running short, high-volume tasks | Claude Haiku 5.5 at low effort | Strong capability at the base rate | See Claude pricing |
| Processing 100K to 272K token documents | GPT-6 Luna | No cliff until 272K | Read the specs |
| Building sub-agents for an Opus planner | Claude Haiku 5.5 | Designed for exactly this role | See models |
| Needing reliable multi-step reasoning | Sonnet 5.5 at medium | Fewer retries per success | See Sonnet 5.5 |
Related on this site: for more context, see GPT-6 Luna, Claude Sonnet 5.5 and AI Pricing Changes.
Frequently Asked Questions About Claude Haiku 5.5
How much does Claude Haiku 5.5 cost?
$0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above that, the whole request is billed at $0.50 and $2.50.
Is it really 90% cheaper than Haiku 4.5?
Per token, yes. But the new tokenizer produces about 30% more tokens for the same text, so the real saving on identical work is closer to 87% in the low tier and less if prompts cross 100,000 tokens.
When was it released?
Anthropic released Claude Haiku 5.5 on October 7, 2026. It will stay available until at least October 7, 2027.
Does it support effort levels?
Yes, it is the first Haiku-class model with adjustable effort. Use low or medium for most tasks and keep high effort for the occasional harder request.
Is it better than GPT-6 Luna?
It scores higher on Artificial Analysis’s index at each effort level, while Luna costs less per task and has a later price cliff. Choose by your prompt lengths and latency needs.
Can it replace Sonnet for coding?
For small edits and test-log triage, often yes. For multi-file changes and hard debugging, Sonnet 5.5 will usually be cheaper per successful task because it needs fewer retries.
Verdict
Move your high-volume, short-prompt work to Claude Haiku 5.5 now. Classification, routing, extraction, compaction and sub-agent chores all get dramatically cheaper, and the speed improvement is noticeable in anything interactive.
The caveat is that the headline 90% is a per-token figure. The tokenizer adds about 30% to your counts, and the 100,000-token cliff can make long prompts five times as expensive as you expected. Benchmark scores here are largely from vendors and a single independent tester.
Your next step: re-measure a day of real traffic with the new tokenizer, add a guard that splits or warns on prompts nearing 100,000 tokens, and only then update the budget forecast.
Sources and Further Reading
- VKTR: Haiku 5.5 pricing, for the tiered rate card.
- Mixed: the tokenizer footnote, for the 30% token-count change.
- The New Stack: Haiku 5.5 launch, for effort controls and benchmark gains.
- Codersera: GPT-6 Luna guide, for the head-to-head index figures.
- Penchan: Haiku 5.5 overview, for subscriber API credits and Claude Code defaults.
On this site: cut spend further with LLM API cost optimization, build agents that browse with AI browser agents, connect tools through the best MCP servers, and compare small models in our GLM-5.3 Flash review.

