Anthropic released Claude Haiku 5.5 on October 7, 2026, calling it its cheapest, fastest and most capable small model so far. It is the first Haiku with effort controls, and input pricing drops to a tenth of Haiku 4.5's for prompts up to 100k tokens. If you run high-volume work (classification, log summaries, triage), it is worth a test.
What changes
According to Anthropic's announcement, Haiku 5.5 costs about 75% less than Haiku 4.5 on average. Prices per million tokens:
| Haiku 5.5 (≤100k) | Haiku 5.5 (>100k) | Haiku 4.5 | |
|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 |
| Output | $0.50 | $2.50 | $5.00 |
| Cache reads | $0.01 | $0.05 | $0.10 |
| Cache writes | $0.125 | $0.625 | $1.25 |
Per the announcement's own footnote, that is 90% cheaper for requests up to 100k tokens and 50% cheaper above that. One detail is easy to miss: prompts over 100k cost five times more per token than short ones, so your savings depend on how much context you send.
The other change is effort control, with Low, Med, High, Xhigh and Max levels shown in Anthropic's charts. It lets you trade cost against quality per task instead of switching models.
On Anthropic's published benchmarks against Haiku 4.5: Terminal-Bench 4.0 goes from 0.0% to 39.2%, OSWorld 2.1 (offline subset) from 15.7% to 72.4%, and Humanity's Last Exam without tools from 10.2% to 45.9%. These are vendor numbers, and Sonnet 5.5 scores higher on every one listed. The page does not state a context window.
On AWS, per the AWS Machine Learning blog post (October 7), it is available on Amazon Bedrock through cross-Region inference profiles (US, EU, AU, JP and Global), in AWS GovCloud (US), and on Claude Platform on AWS in North America. The global profile ID is global.anthropic.claude-haiku-5-5 and it works with the Messages API, Invoke and Converse. Anthropic also lists Google Cloud and Microsoft Azure as availability platforms.
Why it matters in production
A small model is only useful if it is cheap and good enough to replace a larger one for part of your traffic. At these prices, volume tasks (classifying alerts, extracting fields from logs, first-line agent replies) change by an order of magnitude in cost. Effort control also simplifies the architecture: one model with several levels instead of two models with your own routing.
There are caveats. The benchmarks come from the vendor and do not replace your own evaluation set. The AWS post warns that the model may return a thinking block before the text block, so code that reads content[0] will break. And because pricing rises for long prompts, a pipeline that stuffs in lots of context may save less than the headline suggests.
What I would do
- Pick a high-volume task that already has an evaluation and rerun it on Haiku 5.5 at different effort levels.
- Measure real cost per request, not just price per token, watching the 100k threshold.
- Review response parsing: select the text block by type, not by position.
- On AWS, grant only
bedrock:InvokeModelandbedrock:InvokeModelWithResponseStreamto callers, and audit usage with CloudTrail and CloudWatch. - Keep the previous model as a fallback until you have your own data.
Sources
- Anthropic: Claude Haiku 5.5 announcement (October 7, 2026).
- AWS Machine Learning Blog: Introducing Claude Haiku 5.5 on AWS (October 7, 2026).
- AWS Weekly Roundup, October 13, 2026.