Claude Haiku 5.5: ten times cheaper, and it passed our whole coding test
Claude Haiku 5.5 costs a tenth of Haiku 4.5 per token. On our private coding test it got all 186 secret checks right, like Sonnet 5.5, for 0.19 cents per job.
Anthropic has a new small model: Claude Haiku 5.5. It is about ten times cheaper per token than Haiku 4.5, which it replaces. Cheap models usually make more mistakes, so we tested it. On our private coding test, Haiku 5.5 got all 186 secret checks right, just like the much more expensive Claude Sonnet 5.5. Haiku 4.5 got 157. And one coding job cost only 0.19 cents.
What is Claude Haiku 5.5?
Haiku is the smallest and fastest of Anthropic's Claude models. Version 5.5 came out on 7 October 2026. Anthropic calls it "the cheapest, fastest, and most capable small model" it has released, in its announcement.
It is made for many small jobs: summaries, sorting messages into groups, looking things up in a database, or helping a bigger model as a "subagent". A subagent is a helper that a bigger AI sends off to do one small part of a job. Developers can use it through Anthropic's API, and through Amazon, Google Cloud and Microsoft. An API is the connection that programs use to talk to Claude.
What does it cost?
For questions up to 100,000 tokens, Haiku 5.5 costs $0.10 per million tokens read and $0.50 per million written. For longer questions it costs five times as much: $0.50 and $2.50. Anthropic says about 90% of the questions sent to the old Haiku were under 100,000 tokens.
Haiku 4.5 cost $1 and $5. So per token, Haiku 5.5 is ten times cheaper. There is one catch: Haiku 5.5 cuts text into smaller pieces, so the same text counts as about 30% more tokens, Anthropic's developer notes say. Anthropic's own estimate is that running Haiku 5.5 costs about 75% less than Haiku 4.5 on average.
What else is new
Haiku 5.5 is the first Haiku where you can choose how hard it thinks, from low to max. Low is faster and cheaper; max gives better answers. By default, it decides for itself when to think before answering.
It can also read much more at once: up to one million tokens, against 200,000 for Haiku 4.5. And it can write answers of up to 128,000 tokens.
For developers there is a change to watch: the old way of setting a thinking budget now gives an error, and the answer can start with a thinking block. Code that expects the answer in the first block may need a small fix.
Our test: three small coding jobs
We used our private test, the same one as for Claude Fable 5.1. The tests in our public test kit are known by now, so these stay secret. There are three jobs:
- D. Fix a money reader: code that turns "€ 1.234,56" into cents, with several bugs. 28 secret checks.
- E. Split a restaurant bill: shared dishes and a tip, with cents that must add up exactly. 17 checks.
- F. Read a CSV file: a spreadsheet saved as text, where commas, quotes and line breaks can hide inside a field. 17 checks.
For example, one check in job D: "1234,5" must become 123,450 cents, so 1,234 euros and 50 cents. Each model did each job three times, in one request through OpenRouter, a service that gives access to many models, with the default settings.
The results: small, cheap and right
Here is how the four models did over all three jobs:
- Claude Haiku 5.5: 186 of 186. Every try right. About 16 seconds and 0.19 cents per job.
- Claude Sonnet 5.5: 186 of 186, in about 17 seconds, but 2.58 cents per job: more than ten times as much.
- GPT-6 Luna: 184 of 186. It read "1234,5" and "€ 0,5" wrong in one try. About 23 seconds and 0.12 cents per job.
- Claude Haiku 4.5: 157 of 186. It made mistakes on the money job and the CSV job in every try. About 8 seconds and 0.81 cents per job.
So the new Haiku is not only cheaper than the old one, it is also much better. On these jobs it was as good as Sonnet 5.5, for a fraction of the price. It was slower than Haiku 4.5, because it thinks before it answers.
What Haiku 5.5 is not for
Our test uses short, clear jobs. Anthropic itself says that for big, complex coding work, where an AI works step by step in a terminal for a long time, Sonnet 5.5 and Opus 5.5 remain the better choice. On that kind of test, Haiku 5.5 scored 39%, against 71% for Sonnet 5.5, according to Anthropic.
So: use Haiku 5.5 for many small, clear jobs, and as a helper next to a bigger model. For a whole project, a bigger model is still safer.
Sonnet 5.5 got cheaper too
With the launch, Anthropic also cut a price for Sonnet 5.5. Reading from the cache now costs $0.10 per million tokens instead of $0.20. The cache is where tools like Claude Code keep text they have already read, so they can read it again cheaply. Anthropic says this makes Sonnet 5.5 about 20% cheaper on most agent work.
Max and Team subscribers also get a new monthly credit for the API: $100 for Max 5x, $200 for Max 20x and up to $500 for a Team. Our Claude Code pricing article still uses the old cache price; we will update it.
How we checked
On 8 October 2026 we read Anthropic's announcement and its developer notes. The screenshots are our own captures of those pages, cropped only.
We ran the test from 6 to 8 October through OpenRouter, with default settings and three tries per job per model. Haiku 5.5 and GPT-6 Luna ran on 8 October; Haiku 4.5 and Sonnet 5.5 ran earlier for our Fable and Qwen articles. The costs are what OpenRouter charged. The owner of this site paid for the requests. The charts were made by us. None of the images was made by AI.
This was a small test with three short jobs. It says nothing about long projects.