Tech

Anthropic cut the price of its smallest model by 90%

Claude Haiku 5.5 costs 10 cents per million input tokens for most prompts, as the AI price war moves to small, fast models.

Published

1 day ago

October 21, 2026

Massih Shahbazi

Rows of server cabinets in a dark data center, in black and white, beside the headline Ten cents a million

Most AI launches are sold on the word “smarter.” Anthropic’s latest small model was sold on a different word: cheaper. Much cheaper.

Anthropic released Claude Haiku 5.5 on Oct. 7, pricing it at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens, VentureBeat reported. That’s a 90% cut from the $1 and $5 its predecessor, Haiku 4.5, cost. Above 100,000 tokens, the price rises to $0.50 and $2.50, according to The New Stack.

What a token costs, in human terms

A token is a chunk of text, roughly three-quarters of a word in English. A million tokens is something like a long novel or two. At ten cents per million going in, a developer could feed the model an enormous amount of text for less than the price of a cup of coffee.

Most people will never see those prices directly. They matter because they shape what companies build. When reading a document costs a fraction of a cent, it becomes affordable to have AI read every support ticket, every invoice, every log file, every email. Cheaper models don’t just do the same jobs for less. They make new jobs worth automating.

Anthropic estimated that average workload costs would be about 75% lower than with Haiku 4.5, according to Android Authority, a smaller figure than the headline cut because real workloads mix short and long prompts.

A price war, out in the open

VentureBeat’s headline framed the price as matching OpenAI’s GPT-6 Luna. That’s the context that matters most. The biggest AI labs are competing not only on capability but on what it costs to run their models at scale, and small, fast models are where that fight is sharpest. Companies with high-volume tasks will shop on price the way airlines shop for fuel.

Price wars are good for buyers in the short run. They also raise an obvious question about who pays for them. Training and running frontier models is famously expensive, and when prices fall 90% in one release, either the underlying costs fell dramatically or someone is choosing to earn less per query to win market share. Anthropic hasn’t said publicly which it is.

The dial nobody asked for, until they did

Haiku 5.5 is the first Haiku model with effort controls, with medium as the default, according to AI Polix and other coverage. Effort controls let a developer decide how much reasoning the model spends on a task: less for simple sorting jobs, more for harder ones. It’s a way to trade speed and cost against quality on a per-request basis.

Anthropic also added beta computer-use and browser-use support to its Python and TypeScript software kits, according to the launch coverage. Those features let a model operate software or web pages on a user’s behalf. Combined with a very low price, that points to where small models are headed: not just answering questions, but running large numbers of small automated tasks in the background.

Anthropic pitched the model for high-volume work such as summarization, classification and running “subagent” tasks for larger systems, the coverage reported. It’s available through Anthropic’s own API as well as Amazon Bedrock, Google Cloud and Microsoft Foundry, according to StreetInsider.

What the benchmarks don’t prove

Anthropic reported performance gains over Haiku 4.5, but those numbers come from the company, not from independent testing, so treat them as claims until outside evaluations catch up. Coverage also varied in quality: at least one outlet described architectural details that more established publications didn’t confirm.

The other open question is what “cheap enough to use everywhere” means in practice. A model reading every email and every log file is also a model with access to every email and every log file. The cheaper automation gets, the more important it becomes to ask what data is being fed in, where it goes and who’s accountable when an automated agent gets something wrong.

There’s also a disclosure worth making plainly: AIME 360 uses AI tools, including Anthropic’s, in parts of its production work. That doesn’t change the facts reported here, all of which come from the outlets linked, but readers deserve to know it when a story is about a company whose products we use.

From the developer’s desk

For a small startup, a 90% price cut is the difference between a feature that loses money on every user and one that doesn’t. For a big company, it’s a budget line that suddenly shrinks. Either way, it’s the kind of change that quietly gets built into products people use every day without knowing which model is underneath.

That’s the strange thing about the cheapest model in a lineup. It rarely gets the attention the flagship does. But it’s often the one that ends up running the most.