Anthropic released Claude Haiku 5.5 on October 7, the third model in its Claude 5.5 lineup after Opus 5.5 and Sonnet 5.5. It is the smallest and cheapest of the three, and Anthropic positions it for high-volume work and as a subagent beneath larger models.

Pricing and specs

For prompts up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens; longer prompts are billed at $0.50 and $2.50. Haiku 4.5 was priced at $1 and $5, and Anthropic says the new model costs around 75% less to run on average, since the headline cut applies to shorter prompts only. The model has a 1M-token context window and up to 128K output tokens. Its API ID is claude-haiku-5-5, and it is available through Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure.

It is the first Haiku model with adjustable effort. Adaptive thinking is on by default and the effort setting, which defaults to medium, trades reasoning depth against latency and cost. Secondary coverage reports that the API rejects non-default temperature, top_p and top_k values.

Where it fits

Anthropic's framing, as relayed by third-party coverage, is that a larger model such as Sonnet 5.5 or Opus 5.5 plans and judges while Haiku 5.5 handles reading, compaction and narrow subtasks. Figures reported from Anthropic's own benchmark table show 72.4% on the offline subset of OSWorld 2.1 versus 83.9% for Sonnet 5.5, and 39.2% on Terminal-Bench 4.0 versus 70.6% for Sonnet 5.5. Those numbers are vendor-reported, and we have not seen independent verification. They suggest Haiku 5.5 is strong on computer-use tasks for its price but noticeably behind Sonnet on long agentic coding work.

Why it matters

Agent systems often make many model calls per task, so per-token price largely determines what they cost to run. Coverage notes the pricing matches OpenAI's GPT-6 Luna at the same $0.10 and $0.50 rates, pointing to intensifying competition at the low end of the market. Developers should check one open question raised in analysis: Anthropic's documentation does not spell out whether cached tokens count toward the 100K-token pricing threshold.