Anthropic: Claude Haiku 5.5

Claude Haiku 5.5 is Anthropic's October 2026 Haiku-class model, positioned for high-volume, latency-sensitive workloads such as classification, routing, extraction, summarization, compaction, and subagent orchestration alongside larger Claude models. It ships with adaptive thinking and adjustable effort (low through max), a one-million-token context window, up to 128K output tokens, and text-and-image input with text output. API pricing is tiered by prompt length (lower rates up to 100K tokens), with prompt caching and batch discounts available on the Claude Platform and major cloud marketplaces.

Relative to Haiku 4.5, Anthropic reports large gains on agentic coding, tool use, computer use, and multimodal chart understanding, with published scores including 83.7% on SWE-bench Multilingual, 64.8% on SWE-Bench Pro, 57.4% on Humanity's Last Exam with tools, and 1620 Elo on GDPval-AA v2.1 at max effort. The model is not open-weight; parameter count and training details are undisclosed, and access is via the claude-haiku-5-5 API and hosted endpoints rather than local inference.

Haiku 5.5 targets tasks where cost and speed matter more than frontier depth on the hardest multi-hour engineering jobs, while still supporting coding agents, browser and computer use (in beta SDK flows), and multilingual knowledge via strong Global-MMLU performance. Knowledge is current through June 2026 per Anthropic documentation.