Anthropic Releases Claude Haiku 5.5, a Small Model Priced Like a Rounding Error

Anthropic released Claude Haiku 5.5 on Wednesday, and the headline number is a price: $0.10 per million input tokens and $0.50 per million output tokens, for prompts up to 100,000 tokens. Its predecessor, Haiku 4.5, charged $1 and $5. The word for this in economics is “a sale.” The word for it in AI is “Tuesday.”

Some precise facts, because the discount is more nuanced than the banner. Per LetsDataScience, prompts over 100,000 tokens cost $0.50 input and $2.50 output, still half of Haiku 4.5’s rates. Anthropic says the cheaper short-prompt tier covers about 90% of requests to the previous Haiku model. Anthropic describes the model overall as costing “around 75 percent less than Claude Haiku 4.5 for most tasks,” a blended figure. Japanese outlet gori.me notes that the new tokenizer uses slightly more tokens per task, which is baked into the average, and that people who feed it long documents will see smaller savings. In other words: 90% off if you are brief, 50% off if you are not, and about 75% off in the aggregate. The more you talk, the less the discount cares.

The pitch, per AWS, is that Haiku 5.5 is the fastest and most efficient model in the Claude 5.5 family, built for subagents and high-volume, cost-sensitive work. It is the first Haiku with effort controls, meaning a developer can decide how hard it thinks about a given task. For the first time in the history of the small model, you can ask it to try harder. Most of us have been asking that of small things for years, with mixed results.

AWS also describes a division of labor: Claude Opus 5.5 plans and makes judgment calls, and Haiku handles fast, well-defined subagent tasks. This is the org chart of the future: one expensive model with opinions, and a swarm of inexpensive ones with deadlines. If this sounds like a middle-management structure, that is because the training data was full of middle-management structures.

Now the benchmark, which deserves care. LetsDataScience reports that on the offline subset of OSWorld 2.1, a computer-use benchmark, Haiku 5.5 scored 72.4%, versus 15.7% for Haiku 4.5. The article is careful to say these are vendor-reported results that may not hold in every application. We pass along that caution with the enthusiasm it deserves.

Even so. From 15.7% to 72.4% is the kind of improvement normally reserved for people who were told they’d have to retake the exam. A computer-use model that previously operated a computer the way a toddler operates a stand mixer is now, per the vendor, competent in about seven tasks out of ten. The remaining three are presumably the ones involving a CAPTCHA, an unexpected cookie banner, or a printer.

Anthropic also announced beta computer-use and browser-use support in its Python and TypeScript SDKs, per LetsDataScience, which advises evaluating them with appropriate permissions and monitoring, especially where browser actions affect external systems. That is the polite way of saying: the intern can now click things.

There is a pricing story here, too. LetsDataScience reports a cache-read price cut for Claude Sonnet 5.5, from $0.20 to $0.10 per million tokens. When the cheap model gets 90% cheaper and the mid-tier model’s cache gets halved, the effect on the rest of the market is a matter for the rest of the market. Anthropic can set its own prices; the competition can read them and sigh.

For availability, AWS says Haiku 5.5 is on Amazon Bedrock through cross-region inference profiles, in AWS GovCloud (US), and on Claude Platform on AWS in North America. The model ID is claude-haiku-5-5, a name as modest as the price. Not a Sol. Not an Astra. Just a Haiku, which at this point in the naming wars reads as an act of restraint.

The honest reading is that small models keep getting better at the exact things large models used to be needed for, at a fraction of the cost, and that the price of an AI action is heading toward the price of a thought you didn’t have. At $0.10 per million input tokens, you could feed a model the entire text of a long novel several times over for about the cost of a gumball. Whether it should read the novel is a separate question, answered by the effort control.

Try harder, the dial says. Haiku 5.5, for ten cents, will.

Sources: