Anthropic Launches Haiku 5.5: Its Cheapest and Fastest Claude Model Yet

What it claims

Claude Haiku 5.5 (API name claude-haiku-5-5) is Anthropic's new small model, aimed at high-volume work such as document summarization and database querying, and at latency-sensitive work such as live customer support and browser operation. The headline price is a 90% cut, but the realistic saving is smaller. Prompts above 100,000 tokens get only a 50% cut from Haiku 4.5's $1/$5. Haiku 5.5's tokenizer also splits text into slightly more tokens. Anthropic therefore estimates the average saving at about 75%, based on roughly 90% of Haiku 4.5 requests falling under the 100k line. If your workload is long-context heavy, expect less.

Evidence

All figures are Anthropic's, as reported by the page:

Some caveats on reading these numbers:

• OSWorld is reported on an offline subset with partial credit. It is not a full-run strict success rate.
• Terminal-Bench is the clearest capability gap. On agentic terminal tasks, Haiku 5.5 completes roughly half as many tasks as Sonnet 5.5. Haiku 4.5 scored zero on this benchmark.

How to use it

Model ID: claude-haiku-5-5

Where it runs: Claude website, AWS, Google Cloud, and Microsoft Azure.

Pricing:

• $0.10 per million input tokens and $0.50 per million output tokens, for prompts up to 100k tokens.
• Prompts over 100k tokens are priced at 50% of Haiku 4.5's $1/$5.
• Re-estimate your costs with the new tokenizer rather than assuming the 90% figure.

Effort setting: This is the first Haiku with an adjustable effort setting, which trades cost for answer quality. The page does not describe its levels or API parameters.

Related changes:

• Sonnet 5.5's cache-read price is halved to $0.10 per million tokens. This matters if you route harder calls to Sonnet with heavy prompt caching.
• Monthly API credits are rolling out this week: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled across users for Team plans.

Limits

• Failure on a simple test. In Decrypt's single hands-on test, a simple logic question, the model answered almost instantly but got it wrong. This is one anecdote, not a measurement. Still, it is consistent with Anthropic positioning the model for repetitive, easy, non-creative tasks.
• No independent benchmarks. All benchmark numbers come from Anthropic. The page reports no independent evaluation.
• Missing details. The page gives no latency figures, context window, or rate limits, and no details on how the effort setting works.