Anthropic announced Claude Haiku 5.5 on October 7, 2026. The company positions it for repetitive work such as classification and summarization, with an adjustable effort setting. The official model identifier is claude-haiku-5-5. Read the original announcement.
What changed
The launch page lists API input at $0.10 and output at $0.50 per million tokens for prompts up to 100,000 tokens. Above that threshold, the listed rates are $0.50 and $2.50. Cache operations have separate rates. Anthropic says the model is available through its platform and major cloud providers; confirm the model and terms in the platform you use.
What this means for a small application
Our practical reading is that a narrow extraction or sorting task is a useful evaluation target. Build a small set of documents where you already know the answer. Record missing fields, incorrect labels, and outputs that require manual correction. Include messy examples rather than testing only a clean demonstration.
Measure the cost of an accepted result. A cheaper request can still become an expensive task if it requires several retries or a second model to fix the output. Keep prompts, effort settings and output limits consistent while comparing alternatives. Record both response time and total time until your application has a usable answer.
What to check before changing a deployment
- Confirm the exact model identifier in your provider's console.
- Test long and short inputs separately, including cases close to a billing threshold.
- Keep representative failure cases as regression examples.
- Use a limited rollout with a way to return to the current model.
The price figures above describe the announcement's API terms, not a monthly chat membership. Check the current rate card before budgeting. Our subscription and API pricing guide explains the billing distinction, and the Claude profile links official entry points.
This report records the October 7 source announcement; this site's publication date is October 12. We have not run an independent performance benchmark.
