Anthropic releases Claude Haiku 5.5

Anthropic released Claude Haiku 5.5, describing it as the cheapest, fastest and most capable model yet in its small-model tier. Anthropic said the model costs around 75% less to run on average than Haiku 4.5, with the biggest savings on prompts under 100,000 tokens — the range Anthropic said covers about 90% of requests to the previous Haiku model. That brought its headline pricing roughly in line with OpenAI’s GPT-6 Luna, though the two diverge on where surcharge tiers for longer prompts kick in.

Anthropic framed Haiku 5.5 as a supporting worker rather than a standalone flagship: a model meant to summarise documents, classify information, query databases and handle smaller subtasks for larger models such as Opus to assemble, alongside newer workloads like subagent orchestration and autonomous browser use. The company reported gains on agentic benchmarks over its predecessor, including stronger scores on OSWorld 2.1’s offline subset and Terminal-Bench 4.0, Anthropic’s benchmarks for computer-operation and agent-based coding tasks respectively.

The release continues a pattern across frontier labs through 2025 and 2026 of pushing capability down into cheap, fast model tiers rather than only up into ever-larger flagships, on the view that high-volume agentic work is more sensitive to latency and per-token cost than to peak benchmark scores. Haiku 5.5 shipped alongside a cut to the cache-read price of Claude Sonnet 5.5, which Anthropic said lowers the cost of most agent tasks by about a fifth.