Anthropic releases Claude Haiku 5.5 small model and halves Sonnet 5.5 cache read prices

Anthropic has unveiled Claude Haiku 5.5, a new small model priced at about a quarter of Haiku 4.5's cost. Meanwhile, Claude Sonnet 5.5's cache read prices have been halved. Following the Sept. 22 launch of Opus 5.5, this marks three models in the 5.5 generation.

Designed for repetitive tasks, Haiku 5.5 targets high-volume summaries, classification, and coding teams using it as a subagent for smaller tasks. Despite no Anthropic model running faster at standard speed, it's recommended for live customer support and browser automation.

Running costs are significantly reduced, with Haiku 4.5 priced at $1 per million input tokens and $5 per million output. Haiku 5.5 offers 90% cheaper input and output rates at 10 and 50 cents per million tokens respectively, with further discounts on longer prompts. The model also features a new tokenizer that uses slightly more tokens per task, with adjustable effort settings.

Haiku 5.5 outperforms OpenAI's GPT-6 Luna on all six benchmark tests, achieving 72.4% on OSWorld 2.1's offline subset, compared to Luna's 48.9%. For complex agentic coding, Sonnet 5.5 and Opus 5.5 remain the preferred choices.

Asana Inc. found Haiku 5.5 to be 30% quicker in task completion latency and up to 2.5 times faster in inference per agent turn, enhancing the overall user experience. Anthropic reported fewer instances of misaligned behavior in alignment testing, with enhanced cybersecurity safeguards. Access to penetration testing is restricted, but organizations can apply to Anthropic's Cyber Verification Program.

Additionally, Sonnet 5.5's cache read prices have been cut from 20 cents to 10 cents per million tokens, potentially reducing costs by 20% for most agentic work. Claude Max and Team subscribers will receive monthly API credits for the Claude Platform, starting this week. Haiku 5.5 is now available on the Claude Platform, AWS, Google Cloud, and Microsoft Azure.