What changed
Anthropic launched Claude Haiku 5.5 on October 7 as its fastest and most capable small model, aimed at high-volume, latency-sensitive work. The product positioning is more specific than a normal model upgrade: Haiku 5.5 is meant for classification, extraction, routing, compaction, summaries, and subagent turns that would be too expensive to run through a frontier model every time.
The economics are the headline. Anthropic says Haiku 5.5 costs around 75% less to run on average than Haiku 4.5. Its platform model overview lists input pricing from $0.10 per million tokens and output pricing from $0.50 per million tokens for prompts up to 100,000 tokens, alongside a 1M-token context window, 128K maximum output, and adaptive thinking controlled through the effort parameter.
The surrounding developer stack moved with it. The TypeScript SDK v0.132.0 release adds the Haiku 5.5 model and typed browser/computer toolset calls. Anthropic’s SDK toolset documentation describes Python and TypeScript classes that run browser or desktop loops with developer-defined policies and approval callbacks.
Why it matters
Agent products are constrained by economics as much as by intelligence. A model that is good enough for the repetitive middle of a workflow can determine whether an agent is viable at scale. Haiku 5.5 gives builders a clearer worker tier: let the smaller model handle short tool calls, retrieval, routing, context compaction, and parallel subagent work, then escalate ambiguous or high-stakes steps to Sonnet or Opus.
The effort control also turns model selection into an application policy. A product can start with a lower-cost setting for simple, latency-sensitive operations and spend more reasoning budget only when its evaluations show that quality improves. The 1M-token context window is useful for long-running agents, but prompts above 100,000 tokens have higher pricing, so context size still needs to be managed deliberately.
Browser and computer use make the shift more tangible. The SDK supplies the model-facing toolset, while the application owns the browser or desktop driver, URL and file policies, and approval flow. That separation gives teams a practical path to build faster agents without treating every tool call as an unrestricted action.
Builder takeaway
Treat Haiku 5.5 as a default worker, not a universal replacement. Start with measured routing: use it for high-frequency, bounded steps; reserve a larger model for planning and difficult decisions; and track latency, error rate, escalation rate, and cost per completed task. For browser or desktop agents, keep the approval callback and URL/file policy explicit, and test prompt sizes above 100,000 tokens before assuming the long context is economical.