Claude logo via Wikimedia Commons
What Claude Haiku 5.5 is
Claude Haiku 5.5 is a small, fast model from Anthropic, and Anthropic released it on October 7, 2026. The model ID for developers is claude-haiku-5-5. Anthropic describes it as its cheapest, fastest, and most capable small model, which is a claim worth taking in two parts. The cheapest and fastest parts are easy to check against a price sheet. The most capable part is a comparison with older small models, and the benchmark section below shows how large that gap is.
Anthropic sells its Claude models in tiers. The small tier, called Haiku, is built for speed and low cost. The middle tier is Sonnet, and the top tier is Opus. If you read our write-up of Claude Opus 5.5, you have already met the top of the range. Haiku 5.5 sits at the other end. It is the model you pick when you have a lot of requests to run, each one is fairly simple, and the bill matters as much as the answer.
If you only chat with Claude in an app, you may never choose a Haiku model by name. The people who care most are developers and small teams who call the model through an API, and companies that run it behind a product such as a support chat window. This article is written for both groups. It explains the terms as it goes, so you can follow along without a technical background.
The price, line by line
Anthropic prices the API by the million tokens. A token is a small chunk of text, roughly three quarters of an English word. Input tokens are what you send, including your instructions and any documents. Output tokens are what the model writes back. Anthropic uses two price bands, one for prompts up to 100,000 tokens and one for prompts over 100,000 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, those numbers become $0.50 and $2.50. Cache reads cost $0.01 per million tokens in the lower band and $0.05 in the higher band. Cache writes cost $0.125 and $0.625.
Compare that with the older small model. Haiku 4.5 costs $1.00 per million input tokens and $5.00 per million output tokens. That puts the Haiku 5.5 list price at one tenth of Haiku 4.5 for prompts in the lower band. Sonnet 5.5, the middle model, costs $2.00 for input and $10.00 for output, which is twenty times the Haiku 5.5 rate on both lines in the lower band.
The numbers are easier to read in a short list.
- Haiku 5.5: $0.10 input, $0.50 output per million tokens (prompts up to 100k)
- Haiku 4.5: $1.00 input, $5.00 output per million tokens
- Sonnet 5.5: $2.00 input, $10.00 output per million tokens
For a wider view of how these prices stack up against rivals, see our AI model API pricing comparison for October 2026.
Why the average saving is 75 percent and not 90
If the list price is one tenth of Haiku 4.5, you might expect your bill to fall by 90 percent. Anthropic says the average cost is about 75 percent lower than Haiku 4.5. The gap between those two figures comes from a new tokenizer.
A tokenizer is the part of the system that chops your text into tokens before the model reads it. Haiku 5.5 uses a new one, and Anthropic says it uses slightly more tokens per task than before. The same paragraph can come out as a slightly larger token count, so each task consumes a bit more of the cheaper unit. The result is a real saving that is smaller than the price cut alone would suggest.
This is a useful habit to carry to any model switch. Do not multiply your old bill by the ratio of the new price list. Run a sample of your own requests, read the token counts the API reports back, and calculate from there. Anthropic gives the 75 percent figure as an average, so your own workload could land above or below it.
Adjustable effort, a first for Haiku
Haiku 5.5 is the first Haiku-class model with adjustable effort. The settings are Low, Med, High, Xhigh, and Max. Effort controls how much thinking the model puts into a request before it answers. Low gives a quick reply. Max lets the model work longer on a hard problem.
Think of it as a dial you set per request. A classification job that sorts incoming emails into five folders does not need much thought, so Low is a sensible starting point. A request that asks the model to plan several steps and check its own work may justify High or above. Higher effort generally means more time and more tokens, which means more cost, so the dial is also a way to trade money for quality on the requests that need it.
Anthropic has not published, on the page we are working from, a table that shows exactly how the benchmark scores change at each level. That means you should test your own tasks at two or three settings and compare. A simple approach is to start at Low, check whether the answers meet your standard, and only move up where they fail.
Anthropic's benchmark results
Anthropic reports results on five tests that compare Haiku 5.5 with Haiku 4.5 and Sonnet 5.5. These are Anthropic's own results, so treat them as the vendor's claims until outside testers repeat them. Here they are in the order Anthropic lists the three models: Haiku 5.5, Haiku 4.5, then Sonnet 5.5.
- GDPval-AA v2.1, scored as an Elo rating: 1620, 735, and 1840
- OSWorld 2.1 offline subset: 72.4%, 15.7%, and 83.9%
- Humanity's Last Exam without tools: 45.9%, 10.2%, and 56.9%
- Terminal-Bench 4.0: 39.2%, 0.0%, and 70.6%
- Chartography without tools: 46.4%, 6.4%, and 61.6%
The pattern is clear. Haiku 5.5 is far ahead of Haiku 4.5 on every row. On OSWorld, a test of operating a computer, the score rises from 15.7% to 72.4%. On Terminal-Bench, which tests work in a command line, Haiku 4.5 scored 0.0% and Haiku 5.5 scored 39.2%. Those are large jumps for a model at a tenth of the old price.
Sonnet 5.5 still leads Haiku 5.5 on all five rows. The gap is smallest on GDPval-AA and OSWorld, and it is widest on Terminal-Bench, where Sonnet 5.5 scores 70.6% against 39.2%. That row matters if your work involves running commands and fixing things in a terminal, which is why Anthropic points complex coding jobs to its bigger models.
Anthropic's table also lists GPT-6 Luna, OpenAI's small model, on some rows. On GDPval-AA it shows Luna at 1437, below the 1620 for Haiku 5.5. On Terminal-Bench it shows Luna at 16.4%, against 39.2%. If you are weighing small models from both companies, you can read our look at GPT-6.1 Sol for the OpenAI side of the lineup, then compare on your own tasks. A vendor table is a starting point, and it was assembled by the company that wants you to choose its model.
What customers report
The Anthropic page also quotes early customers. These are customer-reported results, and the companies ran them on their own work, so they may not carry over to yours.
Asana reports a latency reduction of more than 30%. Latency is the delay between sending a request and getting the answer, so a 30% cut means people wait noticeably less. HubSpot reports an average of 92.8% on its CRM suite, a set of tests built around customer relationship management tasks. Box reports a result 11 points above Haiku 4.5 at about half the latency.
Each of those three reports lines up with what a small model is for. Asana's number is about speed. Box mixes quality and speed. HubSpot's number measures how well the model handles business software tasks. None of them tells you how the model does on your data, so a short pilot is still the sensible step before a full rollout.
What Haiku 5.5 is good for
Anthropic lists several jobs where it expects Haiku 5.5 to do well. They are summaries, compactions, database queries, classification, subagents that work beside Opus 5.5 and Sonnet 5.5, live customer support, and browser use.
A few of these terms need a plain explanation. Compaction means shrinking a long conversation or work history into a shorter version so that an agent can keep going without running out of space. A subagent is a helper that handles one piece of a larger job. A common setup has a large model plan the work and hand the small, repetitive pieces to a cheaper model. Haiku 5.5 is meant for that helper seat. Browser use means the model operating a web browser to click, read, and fill in pages on your behalf.
Live customer support is a natural fit because the model must answer quickly, and the volume is high. A shop that gets thousands of questions a day cares about both cost per reply and how long the customer waits. Classification is a similar story. Sorting tickets, tagging reviews, or routing messages is repetitive work where a small model that is right most of the time, at a tiny price, beats a large one that costs twenty times as much.
Database queries are another listed use. Here the model turns a plain-language question into a query, or reads query results and explains them. Summaries are the simplest case of all. If you need a short version of a long document, thread, or meeting transcript, a small model is usually enough.
Anthropic is direct about the limit. It says Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding. If your task is a large software change that spans many files and needs careful planning, the Terminal-Bench gap above is a reason to start with a bigger model.
Where you can use it
Haiku 5.5 is available on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. That matters for companies that already buy their cloud services from one of those providers, because they can use the model within an account and billing setup they already have.
For people who use Claude through the apps rather than the API, Anthropic lists consumer plans at claude.com/pricing. The Free plan costs $0. Pro costs $20 per month, or $17 per month when billed annually, which is $200 paid up front. Max starts at $100 and comes in two sizes, 5x or 20x the usage of Pro. The page we reviewed does not say which model each plan uses by default, so check the app for the model picker if you want to know what you are talking to.
Safety and what the safeguards block
Anthropic says Haiku 5.5 shows fewer misaligned behaviors than Haiku 4.5. Misaligned behavior means the model doing things its builders did not intend, such as ignoring instructions or acting against what the user wanted. Anthropic does not give a count on its page, so the claim is directional.
The company also describes its safeguards in two areas. For cybersecurity, the safeguards are stricter than those on Haiku 4.5 and looser than those on Sonnet 5.5. They still block penetration testing style attacker techniques. For biology, the safeguards match those on Sonnet 5, Sonnet 5.5, and Opus 5.
In practice, a security professional who wants help with legitimate defensive work may hit refusals on some requests. Anthropic runs verification programs for life sciences and cyber work, which exist so that qualified people can apply for access suited to their field. If your job falls in either area, look at those programs before you build a workflow around the model.
Other news from the same announcement
Anthropic bundled a few other changes into the same release. Sonnet 5.5 cache reads were cut by 50%, from $0.20 to $0.10 per million tokens. Subscribers on the Max 5x plan get $100 in monthly API credits, and Max 20x subscribers get $200. The company's SDKs also add beta support for computer use and browser use.
Caching deserves a short explanation because it affects the bill more than most people expect. When you send the same text again, such as a long instruction block or a policy document, the provider can store it and charge a low rate to read it back. For Haiku 5.5 a cache read costs $0.01 per million tokens in the lower band, a tenth of the plain input price. A support bot that resends the same company guidelines with every message is the textbook case. Place the stable text at the start of your prompt and the changing text at the end, so more of each request can be served from the cache.
The Max credits are a practical perk for developers. If you pay for Max to use Claude in the apps, the monthly API credit lets you test things like Haiku 5.5 in a project without opening a separate budget. The page does not describe the fine print, so read the terms in your account.
A worked cost example
Take a job that sends 1 million input tokens and receives 200,000 output tokens. Assume every prompt is under 100,000 tokens, so the lower price band applies, and assume nothing is cached.
On Haiku 5.5, the input costs $0.10 and the output costs $0.10, since 200,000 tokens is 0.2 million and the output rate is $0.50 per million. The total is $0.20. On Haiku 4.5, the input costs $1.00 and the output costs $1.00, for a total of $2.00. On Sonnet 5.5, the input costs $2.00 and the output costs $2.00, for a total of $4.00.
Now scale it to a busy month. If you run that job 5,000 times, Haiku 5.5 costs $1,000, Haiku 4.5 costs $10,000, and Sonnet 5.5 costs $20,000. The new tokenizer would push the Haiku 5.5 figure up a little, since it uses slightly more tokens per task, and that is the reason Anthropic quotes about 75 percent savings and not the full 90 percent from the price list.
Caching changes the picture again. If 800,000 of the input tokens in that job were served from the cache at $0.01 per million, that portion costs $0.008 instead of $0.08. Cache writes, at $0.125 per million, apply when you first store the text, so a prompt reused many times gains the most.
Caveats worth keeping in mind
The first caveat is the source of the numbers. Every benchmark here comes from Anthropic, and every customer figure comes from the customer. Neither group is an independent test lab. They are useful signals, and they are also the best-case story a company chooses to publish.
The second caveat is the size of the model. Haiku 5.5 is strong for a small model, and the Terminal-Bench gap to Sonnet 5.5 shows where it falls short. It scores 39.2% on that test where Sonnet 5.5 scores 70.6%. A cheap model that needs three tries to finish a job can cost more in time and review than a pricier model that finishes in one.
The third caveat is that prices and plan details change. The figures in this article come from Anthropic's launch page as of October 18, 2026. Confirm the current rates before you set a budget. The fourth is the safeguard behavior. If your work touches security testing or biology, plan for refusals and look into the verification programs early.
A sensible way to start is a small pilot. Pick 30 to 50 real requests from your work, run them on Haiku 5.5 at a low effort setting, and mark each answer as acceptable or not. Run the same set on Sonnet 5.5. If the small model passes most of them, route that kind of work to it and keep the bigger model for the hard cases. Many teams split traffic this way, and the savings are large when most of the volume is simple.
Frequently asked questions
When was Claude Haiku 5.5 released?
Anthropic released Claude Haiku 5.5 on October 7, 2026. The model ID is claude-haiku-5-5, and Anthropic calls it its cheapest, fastest, and most capable small model.
How much does Claude Haiku 5.5 cost?
For prompts up to 100,000 tokens, it costs $0.10 per million input tokens and $0.50 per million output tokens. Above 100,000 tokens, it costs $0.50 input and $2.50 output. Haiku 4.5 costs $1.00 and $5.00, and Sonnet 5.5 costs $2.00 and $10.00. Anthropic says the average cost is about 75% lower than Haiku 4.5 because the new tokenizer uses slightly more tokens per task.
What does adjustable effort mean?
Haiku 5.5 is the first Haiku-class model with adjustable effort, with settings of Low, Med, High, Xhigh, and Max. Lower settings give quicker, cheaper answers, and higher settings give the model more room to work on hard problems.
Is Haiku 5.5 as good as Sonnet 5.5?
No. In Anthropic's own results, Sonnet 5.5 scores higher on all five reported benchmarks, such as 70.6% against 39.2% on Terminal-Bench 4.0. Anthropic says Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding. Haiku 5.5 is the cheaper choice for summaries, classification, support, and similar jobs.
Where can I use Claude Haiku 5.5?
It is available on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure.
What are the Claude consumer plan prices?
Free costs $0. Pro costs $20 per month, or $17 per month billed annually with $200 paid up front. Max starts at $100 and offers 5x or 20x the usage of Pro.