Claude Haiku 5.5 Cuts Prices Up To 90 Percent: What The Cheap Tier Of AI Now Does
On 7 October, Anthropic released Haiku 5.5 and cut the price of its small model substantially. On average it now costs around 75 percent less to run than the model it replaces, and for the short requests that make up most of the traffic, the reduction reaches 90 percent. The same week, OpenAI cut the prices of its own mid and small tier models by half.
Two vendors cutting prices in the same month is not a coincidence. It is the shape of the market now: the frontier gets the headlines, and the cheap tier is where the products get built. Here is what Haiku 5.5 actually costs, what Anthropic says it can do, and where the numbers stop being useful.
The pricing table, and the threshold that decides your bill
The first thing to understand about Haiku 5.5 is that its price depends on how long your prompt is.
| Per 1 million tokens | Haiku 5.5, prompts up to 100k | Haiku 5.5, prompts over 100k | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|---|
| Input tokens | 0.10 dollars | 0.50 dollars | 1.00 dollars | 2.00 dollars |
| Output tokens | 0.50 dollars | 2.50 dollars | 5.00 dollars | 10.00 dollars |
| Cache reads | 0.01 / 0.05 dollars | 0.01 / 0.05 dollars | 0.10 dollars | 0.10 dollars |
| Cache writes | 0.125 / 0.625 dollars | 0.125 / 0.625 dollars | 1.25 dollars | 2.50 dollars |
Read those columns again, because the trap is in the middle. Haiku 5.5 is ten times cheaper than its predecessor on input and output for prompts under 100,000 tokens, and it is roughly twice as cheap above that threshold, where the rate jumps five times. Anthropic notes that prompts under 100,000 tokens made up around 90 percent of the requests to the previous Haiku model, which is why the headline discount is framed around that limit.
The practical consequence is that a long document stuffed into Haiku 5.5 can cost more than the same work sent in shorter pieces. Anyone building a pipeline on Haiku 5.5 should watch the token count of the context they attach, not just the prompt they type.
Cache reads are the other number worth noting. At one cent per million tokens for short prompts, cached context costs a tenth of what the input rate charges, which is the same lever every vendor has been pulling since agents started resending the same context on every turn. We covered the mechanics of that in our guide to inference caching, and it is the single most reliable way to cut an AI bill.
What Anthropic says the model can do
Anthropic describes Haiku 5.5 as its cheapest, fastest and most capable small model, aimed at high volume, cost sensitive work: summaries, compactions, database queries and classification, plus use as a subagent on coding tasks and for speed sensitive work such as live customer support and browser use.
The benchmark table it published is a comparison against its own predecessor, against Sonnet 5.5, and against OpenAI's GPT-6 Luna. The numbers, as published by Anthropic:
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| GDPval-AA v2.1, knowledge work | 1620 | 735 | 1437 | 1840 |
| AA-Briefcase v1.1, knowledge work | 1578 | 614 | 1336 | 1824 |
| OSWorld 2.1 offline, computer use | 72.4 percent | 15.7 percent | 48.9 percent | 83.9 percent |
| Humanity's Last Exam, no tools | 45.9 percent | 10.2 percent | not reported | 56.9 percent |
| Terminal-Bench 4.0, agentic coding | 39.2 percent | 0.0 percent | 16.4 percent | 70.6 percent |
| FrontierCode 1.1, agentic coding | 46.4 percent | not reported | 42.4 percent | 52.1 percent |
| Chartography, visual reasoning | 46.4 percent | 6.4 percent | 29.1 percent | 61.6 percent |
Those are vendor numbers from a vendor run evaluation, and the comparison against a competitor's model is drawn from that competitor's published figures rather than from a neutral test. Treat the direction as informative and the decimals as marketing.
What the direction says is that a model at this price is no longer useless for the things small models used to fail at. Computer use at 72 percent on OSWorld, against 15.7 percent for the previous generation, is the most striking jump in the table, and Terminal-Bench going from zero to 39 percent is the difference between a toy and a tool for narrow coding work.
Anthropic also shipped an adjustable effort setting with this model, the first time a Haiku class model has one. That matters more than it sounds, because it turns the model into a cost dial: low effort for classification, higher effort when the task is worth it, on the same model and the same integration.
The customer evidence in the announcement is thin but specific: Asana reported over 30 percent lower latency on task completions and up to 2.5 times faster inference per agent turn compared with the model it was using.
A small model now supports computer use, which is the real headline
Buried in the availability notes is the detail with the longest tail. Anthropic updated its Python and TypeScript SDKs to support computer use and browser use in beta, and it points out that Haiku 5.5 is well suited to those tasks because of its combination of speed, capability and price.
Put plainly: the class of model you would use to drive a browser or click through an interface is now the cheap one, not the expensive one. That changes the economics of agentic products, because an agent that browses and clicks makes a lot of model calls per task. Running those calls on a small model at 10 cents per million input tokens rather than a flagship at 2 dollars is the difference between a product with a margin and one without. Our earlier guide to building production grade agentic systems makes the same argument from the architecture side: the routing decision is where the money is.
Safeguards are tighter than capability here in one specific area. Anthropic says Haiku 5.5's cybersecurity safeguards are more restrictive than Haiku 4.5's, permitting defensive work but blocking penetration testing and techniques more likely to be used by attackers, and that its biology safeguards match those of the larger models in the current line. Researchers who need wider access can apply to Anthropic's verification programmes, which is the pattern the whole industry has settled on for this kind of capability.
The capability jump that matters most
The table hides a threshold that every builder has felt. Small models have been cheap and fast for years, and the reason products were still built on expensive ones was never price. It was that a cheap model could not reliably do the things an agent needs to do: click the right control, produce a diff that applies, follow a multi-step instruction without losing the thread.
Haiku 5.5 moves on exactly those axes. Computer use at 72.4 percent on OSWorld 2.1 offline against 15.7 percent for Haiku 4.5 is a near fivefold improvement on a task small models previously failed outright. Agentic coding on Terminal-Bench 4.0 going from 0.0 to 39.2 percent is the difference between a model that cannot finish and one that finishes a third of attempts, which for narrowly scoped work is enough. FrontierCode at 46.4 percent, ahead of the 42.4 percent Anthropic reports for GPT-6 Luna, tells the same story.
The pattern is that the cheap tier has crossed the line where it can be trusted with a single well-defined step inside a larger system. That is a smaller claim than saying small models rival large ones, and it is the one the numbers actually support. It is also the claim that matters commercially, because most agent architectures are a chain of simple steps with a few hard decisions in the middle.
The other half of the market moved in the same direction
On the OpenAI side of the aisle, the same month brought GPT-6 Sol and Luna with prices cut by 50 percent against their predecessor's promotional pricing, and a caching overhaul that is familiar reading.
OpenAI says it improved prompt caching for GPT-6 to deliver higher cache hit rates by default, with 90 percent discounts on cached input token reads, plus new tooling: a dashboard showing how much input is cached over time, a diagnostics view explaining missed caching opportunities, and explicit cache breakpoints so developers choose where cached prefixes end. It reports that GitHub measured a reduction of more than 50 percent in the share of prompt tokens requiring fresh processing across billions of requests.
OpenAI's benchmark claims are aimed squarely at Anthropic's models, as Anthropic's are aimed at its. On AutomationBench, OpenAI says Sol at its highest effort setting outperforms Claude Opus 5 at maximum effort at 9 percent of the cost per task, and that Luna improves on its predecessor by 5.4 percentage points at 58 percent lower cost. On DeepSWE v1.1, it reports Luna at 66.6 percent, which it describes as comparable to Claude Opus 5 and Claude Fable 5 at medium effort, at 93 to 96 percent lower cost per task.
Both companies are publishing comparisons that flatter themselves, using benchmarks they chose and competitor figures taken from public reports. The useful signal is not who wins each row. It is that both are now competing primarily on cost per completed task rather than on raw capability, and that both have concluded caching is the mechanism.
What this changes if you are the one paying the bill
Four practical shifts fall out of the pricing tables.
Routing gets cheaper to justify. The gap between the flagship tier and the small tier is now wide enough that splitting work is obviously worth the engineering. Classification, extraction, summarisation and compaction belong on Haiku 5.5 or its equivalent, and only the reasoning steps need anything more expensive. Our earlier piece on the Gemini 4 Argon pricing changes worked the same arithmetic from the flagship side.
Subagents become the default pattern. When a Haiku 5.5 subagent costs a twentieth of the main agent, the design question flips from whether you can afford to spawn one to how many you should.
The trial economics changed for subscribers. Anthropic is giving Max and Team subscribers a monthly API credit, 100 dollars for Max 5x users, 200 dollars for Max 20x, and up to 500 dollars pooled for Team plans, usable on any model. For a small team experimenting with agents, that is a meaningful amount of free runway, and it is the kind of thing that decides whether a developer builds on one platform or another.
Caching work pays for itself faster. With cache reads at a cent per million tokens on short prompts, the engineering habit of keeping stable prompt prefixes byte identical has a clearer return than it did a year ago. If your prompts reshuffle retrieved documents on every call, you are paying full price for content you already sent. The broader shift this is part of, from renting intelligence to running it cheaply at volume, is the same trend we described in why small language models are quietly winning.
There is a competitive angle worth noting here too. A model at these prices makes the cheapest tier the default place to experiment, and any team that held back on an agentic feature because the token bill looked unpredictable now has a version of it they can trial on Haiku 5.5 and promote to a larger model only where testing says the extra spending is justified.
Six things worth building on a model this cheap
- Compaction inside long agent runs. When a context window fills, something has to summarise what came before. Haiku 5.5 makes that step cheap enough to run often rather than once in desperation.
- Classification and routing. Deciding which request deserves the expensive model is a classification problem, and this is exactly the price point for high volume decisions.
- Subagents. Sending several workers to search, extract or check work is only affordable when the workers are cheap, and at 10 cents per million input tokens they are.
- Browser and computer use. With SDK support in beta and prices at this level, the class of agent that clicks through interfaces moves from research demo to product decision.
- Document extraction at volume. Invoices, contracts and forms are work where the answer is checkable and the volume is high, which is the ideal shape for Haiku 5.5.
- Evaluation and regression testing. Running your own test sets against a cheap model on every deploy is how you find out that a prompt change broke something, and this release makes that routine rather than a monthly exercise.
Where the numbers stop
Four limits on everything above.
- Every benchmark in this article is vendor reported. Anthropic's table is Anthropic's evaluation, OpenAI's table is OpenAI's evaluation, and neither ran a neutral comparison of the other. I have not tested either model.
- The 90 percent figure applies below 100,000 tokens. Above that threshold Haiku 5.5's rate jumps five times, so the headline discount does not describe long context work.
- Cost per task comparisons are affected by retries, tool calls and how the agent is designed. A cheap model that needs three attempts can cost more than an expensive model that needs one, and neither vendor's table prices in your own implementation.
- The adjustable effort setting and the computer use SDK support are new, so expect Haiku 5.5's behaviour at each effort level to be retuned over the coming months.
FAQ
Is Claude Haiku 5.5 cheaper than Haiku 4.5?
Yes, substantially, and the size of the reduction depends on your prompt length. For prompts under 100,000 tokens, input and output tokens cost a tenth of the previous model's rates. Above that threshold, the discount is closer to half.
What is Haiku 5.5 best used for?
Anthropic's own guidance points at high volume, cost sensitive work: summarisation, compaction, classification, database queries, and subagent roles inside larger agentic systems, plus speed sensitive work such as live support and browser use. It says explicitly that Sonnet and Opus remain better choices for complex agentic coding.
Where can I use it?
Haiku 5.5 is available on the Claude Platform with the model identifier claude-haiku-5-5, and through Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic also added computer use and browser use support to its Python and TypeScript SDKs in beta alongside this release.
Does Haiku 5.5 make a larger model unnecessary?
For narrow tasks, often yes. For multi-step reasoning, no. Anthropic's own table puts Sonnet 5.5 well ahead on the benchmarks that require sustained reasoning and complex coding, and its guidance points that work at Sonnet and Opus. Haiku 5.5 handles the many small steps, and a larger model gets called for the few decisions that determine whether the result is right. If you take one design lesson from this release, take the routing one.
The short version
Haiku 5.5 arrives at 10 cents per million input tokens and 50 cents per million output tokens for prompts under 100,000 tokens, with cache reads at a cent, an adjustable effort setting, and a benchmark table that shows large gains over the previous generation on computer use and agentic coding. OpenAI cut its own small model prices by half the same month, and both vendors now compete on caching and cost per task rather than raw capability. The practical consequence is that routing, subagents and cache design have all become cheaper to justify than they were a month ago, and the trap to avoid is the fivefold price jump above 100,000 tokens.
Sources
- Anthropic, Introducing Claude Haiku 5.5, including the full pricing table, benchmarks, safeguards and availability: https://www.anthropic.com/claude-haiku-5-5
- OpenAI, introducing GPT-6 Sol and Luna, including the 50 percent price reduction and the caching changes: https://openai.com/index/introducing-gpt-6-sol-and-luna