Anthropic has released Claude Haiku 5.5, its smallest and cheapest model in the Claude 5.5 family, targeting businesses that need AI to handle a lot of relatively quick tasks without running up a large computing bill.
Haiku 5.5 arrives after Claude Opus 5.5 and Sonnet 5.5, completing Anthropic’s latest three-model lineup. The company says its newest Haiku is not only cheaper than the previous generation but also considerably more capable, particularly when used for summaries, classification, database queries, browser work and customer support.
For companies making thousands or even millions of AI calls, though, the pricing may end up being the most important part of the release.
Haiku 5.5 Is Built for Jobs That Happen at Scale
Not every AI request needs a company’s biggest model.
A bank sorting support tickets, an online store generating short product summaries or a software company running background AI agents may care more about speed and cost than getting the strongest reasoning model available.
That is where Claude Haiku 5.5 is meant to fit.
Anthropic describes it as suitable for repetitive, high-volume work such as summarisation, extracting information from documents, compacting context and classifying data. It can also operate as a smaller subagent while a more powerful Claude model handles the difficult part of a larger job.
That setup can save money. Instead of asking an expensive model to carry out every small step, businesses can send routine work to Haiku and reserve Sonnet or Opus for situations that actually require them.
The Price Difference Is Hard to Ignore
Anthropic says Haiku 5.5 costs around 75% less to run on average than Haiku 4.5.
For prompts of up to 100,000 tokens, input pricing starts at $0.10 per million tokens and output pricing at $0.50 per million. Longer prompts cost more, but the rates remain well below the company’s larger models.
That matters because AI costs look very different once usage reaches enterprise scale.
A few chatbot questions do not cost much. Millions of background requests, customer conversations and document checks do.
The lower Claude Haiku 5.5 pricing makes it easier to use AI in situations where the economics may previously have been difficult to justify.
Anthropic says roughly 90% of requests to its previous Haiku model used prompts below the 100,000-token level, meaning the cheapest pricing tier should cover much of the workload the model is designed for.
It Is Also Anthropic’s Fastest Standard Model
Cost is only half of the pitch.
Anthropic claude calls Haiku 5.5 its fastest model at standard speed. That distinction matters for applications where the person on the other side is waiting for an answer.
A support chatbot that pauses for too long feels broken even if the eventual response is good. The same applies to voice assistants, browser agents and AI built directly into software.
Early testing shared by Anthropic points to shorter response times in several business applications.
Asana, for instance, reported lower latency in its internal tests, while other companies evaluated the model on document analysis, CRM work and high-volume enterprise tasks.
Those are exactly the situations where shaving a small amount of time from every request can become noticeable.
Haiku Is Getting More Capable, Not Just Cheaper
Smaller AI models used to involve a fairly obvious compromise. They were faster and cheaper, but the drop in capability could be large.
Anthropic is trying to narrow that gap.
Its own benchmark results show substantial gains over Haiku 4.5 across computer use, professional knowledge work, coding and visual reasoning.
On the OSWorld computer-use evaluation, for example, Haiku 5.5 scored far above its predecessor. It also showed a large improvement on Anthropic’s agentic coding tests.
Benchmarks do not tell the whole story, particularly when businesses have very specific workloads, but the direction is notable.
The Anthropic Haiku model is becoming something companies may use for actual agentic work rather than merely simple text generation.
Developers Can Now Choose How Much Effort It Uses
Haiku 5.5 also gets an adjustable effort setting for the first time in the Haiku family.
Developers can decide whether a task should prioritise lower cost or allow the model to spend more effort producing a stronger answer.
It gives businesses another way to control spending.
A straightforward classification request may not need much reasoning. A more complicated task can be allowed additional effort without automatically being handed over to a larger model.
That kind of flexibility is becoming increasingly important as AI systems move into production environments where thousands of different tasks may run through the same application.
Anthropic Has Tightened Safety Controls Too
The new model comes with revised safeguards alongside the performance improvements.
Anthropic says Haiku 5.5 showed fewer instances of misaligned behaviour during its evaluations and was less willing than Haiku 4.5 to cooperate with harmful requests.
Cybersecurity restrictions have also changed.
The model allows a wider range of defensive security work than some of Anthropic’s larger recent models, while still blocking techniques the company considers more likely to be used offensively.
Its biology safeguards follow the same general structure used across other recent Claude models.
For businesses, these details become more relevant as smaller models are trusted with autonomous or semi-autonomous tasks rather than being used only to answer questions.
Haiku 5.5 Can Work as Part of a Bigger AI Team
One of the more interesting uses Anthropic highlights is pairing Haiku with Sonnet 5.5 or Opus 5.5.
Imagine a larger model preparing a financial report.
Instead of reading every filing itself, it could send a smaller Claude Haiku 5.5 agent into individual documents to pull out specific numbers. Haiku handles the repetitive lookup; the larger model handles the interpretation and final output.
That kind of model routing is likely to become more common.
The best AI system may not be one giant model doing everything. It may be a group of models of different sizes, with each one used only where its cost and capability make sense.
For Anthropic, having Haiku, Sonnet and Opus in the same generation makes that easier to build.
Claude Sonnet 5.5 Is Getting Cheaper as Well
Anthropic used the Haiku announcement to make another pricing change.
Cache reads for Claude Sonnet 5.5 have been cut in half, from $0.20 to $0.10 per million tokens.
The company estimates that this can reduce the cost of many agent-based Sonnet workloads by around 20%.
It is another sign of how quickly competition in AI is moving away from raw intelligence alone.
Companies now want models that are capable, but they also care about latency, token use and the cost of running agents continuously.
Haiku 5.5 is Anthropic’s clearest response to that pressure yet.
Businesses Can Start Using It Immediately
There is no waiting period for developers.
Claude Haiku 5.5 is available through Anthropic’s own platform as well as Amazon Web Services, Google Cloud and Microsoft Azure.
For Anthropic, the timing is useful. Businesses are experimenting with more autonomous AI systems, but running a large model for every small task remains expensive.
Haiku 5.5 gives the company a model aimed directly at that gap.
It does not replace Sonnet or Opus, and Anthropic itself says those remain better choices for difficult agentic coding and more complex work.
The point is different.
When a company needs an AI model to do something simple thousands of times a day, speed and economics start to matter almost as much as intelligence. Haiku 5.5 is built for exactly that job.
