Claude Haiku 5.5 API pricing and routine agent tasks

Claude Haiku 5.5 launched on October 7: $0.10 per 1 million input tokens and $0.50 per 1 million output tokens, compared with $1 and $5 for Haiku 4.5. An agent using 10 million input and 2 million output tokens a month used to cost $20; now it costs $2, provided the requests are shorter than 100,000 tokens. Anthropic puts the average saving at 75%.

Give it routine agent work: sorting incoming requests or summarizing a customer conversation for a manager. A powerful model should, of course, do the thinking; a simpler one can handle execution.

In the Hermes example described here, easier tasks are assigned to less capable models, and the most demanding tasks go to the smartest ones. Token savings in that example reached 80%.

Claude Haiku 5.5 API pricing: what launched on October 7#

If you use Claude Code through the Anthropic API, you do not need to switch anything: since v2.1.293, released that same day, the default Haiku model there has been Haiku 5.5. Everything that ran on Haiku is already running on a model that costs up to 10 times less per token.

The claude-haiku-5-5 model can read up to 1 million tokens at a time and produce a response of up to 128,000 tokens. For the first time, Haiku has an effort setting: how hard the model thinks. It is available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.

The savings apply to requests of up to 100,000 tokens, which Anthropic says account for around 90% of requests to the previous Haiku. For longer requests, the price is 5 times higher.

Those staying with Sonnet 5.5 also pay less: cache reads have halved in price, from $0.20 to $0.10 per 1 million tokens, and Anthropic says agentic tasks cost about 20% less.

What this changes for your agent and how to try it#

Anthropic itself recommends the model for this kind of routine work and as an assistant to Opus 5.5 and Sonnet 5.5. You can give simple routine work to cheaper models, and you should. You would not send a company director out to buy bread.

Anthropic recommends keeping complex agent tasks with Sonnet and Opus. In Terminal-Bench 4.0, where the model solves multistep command-line tasks, Haiku 5.5 scores 39.2%, versus 70.6% for Sonnet 5.5. In OSWorld, where the model operates a computer itself, the improvement is substantial: 72.4%, compared with 15.7% for Haiku 4.5, though it still falls short of Sonnet 5.5's 83.9%.

How to try it

1. Choose 1 routine step, such as sorting incoming requests. Set its model to claude-haiku-5-5 and leave the other steps alone.
2. If your code was written for Haiku 4.5, remove the manual thinking setting budget_tokens: on 5.5 it produces a 400 error. Adaptive thinking is enabled by default in the new model.
3. Start with the default effort setting, medium. Lower it only if the quality holds up.
4. Run 10–20 of the same tasks that the previous model handled through Haiku 5.5. Compare answer quality—whether a request went missing or an important detail was left out of a summary—and the bill for those tasks in your Anthropic dashboard.

There is a catch. Haiku 5.5 uses new token counting (see Anthropic's release notes), so the same text may count as more tokens and the actual saving may be less than 75%.

You can also try it for free. Over the course of a week, Anthropic will give Max and Team subscribers a monthly API credit: $100 for Max 5x, $200 for Max 20x, and up to $500 shared across a Team, usable on any model. You can claim it in Settings → Billing; unused credit does not roll over.

A practical approach: a simpler model or a way to pay less#

The approach described here for earlier Haiku versions was straightforward: it is a very simple model for simple tasks; do not use it for development. Choose the strongest model for the main work rather than switching to Sonnet to save money: the result matters. Routine agent work in Hermes goes to free or very cheap models.

Likewise, give Haiku 5.5 only the incoming-request sorting step: connect it alongside a powerful model and tell the agent to use it for the easiest tasks.

The order is: start with the most expensive options, then bring the cost down. Get a result with a powerful model, then move steps to Haiku 5.5 one by one and watch for a drop in quality.

Regularly switching models does not, by itself, improve task performance; sometimes it does the opposite. In this approach, the model you have is not the most important thing. Your instructions are.

Which routine task would you move to Haiku 5.5 first? Compare that step with your agent's current result.

Additional material is available on Ailoot.