Anthropic launches Claude Haiku 5.5 on OpenRouter with adjustable reasoning at 75% lower cost

Anthropic has released Claude Haiku 5.5 on OpenRouter, marking the first time the Haiku line gets a reasoning mechanism with tunable effort levels. The smallest model in the family targets high-volume, cost-sensitive workloads — text summarization, subagents, browser use — and replaces Haiku 4.5 with measurable gains in coding, computer use and knowledge tasks, while keeping the 1-million-token context window and joint text-and-image support.

What’s new

OpenRouter’s measurements put Haiku 5.5 above 100 tokens per second at roughly a quarter of the previous generation’s price. The cost drop comes alongside a jump on Anthropic’s internal benchmarks, showing stronger code writing, tool use and long action-chain construction — exactly the scenarios Haiku was built for. The 1-million-token context window remains unchanged, so no trade-off on history length or input size.

Adjustable effort

The headline feature is a three-level “effort” knob — low, medium, high — that controls reasoning depth before the model answers. Thinking is enabled adaptively by default and can be turned off entirely when latency matters more than depth. The logic is straightforward: spend extra tokens and time only when the task justifies it; for simple requests the model responds instantly with no overhead.

Availability and infrastructure

OpenRouter serves the model through multiple cloud providers simultaneously, with automatic routing to a healthy provider when one fails, provided the request settings allow it. Per-provider uptime data is exposed via the Endpoints API, letting production teams build their own fallback logic or track SLAs in real time.