Anthropic has released Claude Haiku 5.5, a small AI model the company says is its cheapest, quickest and most capable in that size class. It is the last of three Claude 5.5 models to arrive, following Opus 5.5 on September 22 and Sonnet 5.5 on September 28. Haiku 5.5 launched on October 9.
Pricing and intended use
Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Haiku 4.5 was priced at $1.00 and $5.00 for the same amounts, so the listed rates are one tenth of the earlier model's.
Longer prompts get a further 50% reduction. Anthropic says about 90% of earlier requests were under the 100,000-token mark. The new model breaks text into slightly more tokens, so the company estimates average savings of roughly 75%.
The model is aimed at repetitive, simple, non-creative work where speed and low cost matter, such as:
- Live customer support chats
- Summarizing long emails and documents
- Querying databases
- Operating a web browser
Benchmark results
| Test | Haiku 5.5 | GPT Luna | Other |
|---|---|---|---|
| OSWorld 2.1 | 72.4% | 48.9% | - |
| Terminal-Bench 4.0 | 39.2% | 16.4% | Sonnet 5.5: 70.6%; Haiku 4.5: 0% |
| GDPval-AA v2.1 (Elo) | 1620 | 1437 | Haiku 4.5: 735 |
These figures show Haiku 5.5 ahead of GPT Luna on computer-use and command-line tasks. The source also cautions that the model can still give wrong answers, even though it works quickly.
Other changes
- Haiku 5.5 adds an adjustable effort setting so users can trade cost against answer quality.
- Anthropic lowered the cache-read price for Sonnet 5.5 to $0.10 per million tokens of text it has already processed.
The releases follow public remarks by Anthropic's CEO in favor of a careful approach to advancing AI capabilities.
Why it matters
Many crypto tools, including chatbots, research assistants and software that reads documents or runs tasks in a browser, depend on AI models billed by the token. A tenfold drop in listed price for a small model shows how quickly the cost of running such services can change. For learners, the release is also a reminder to read benchmark numbers as vendor-reported comparisons and to check AI output, since fast and cheap models can still be wrong.




