Claude Haiku 5.5

• 📏 1M-token context, up from 200K on Haiku 4.5
• 🆕 Max output doubled to 128K tokens
• 🧠 First Haiku with adaptive thinking and effort controls
• ⚡ Built for latency-sensitive routing, classification, extraction, subagents
• 👁️ Handles vision, documents, tool calling, web search
• 🔧 Batches API supports up to 300K output tokens via beta
• 🏢 Anthropic model, released October 2026
• 📚 Newer tokenizer counts ~30% more tokens than 4.5

Anthropic

Anthropic PBC is an American artificial intelligence company headquartered in San Francisco. Structured as a public benefit corporation, the lab develops large language models under the Claude name, with a research emphasis on building reliable, steerable,…

Claude Haiku 5.5 is Anthropic's small, fast tier in the Claude 5.5 generation, released in October 2026 and positioned for high-volume, latency-sensitive work: classification, routing, extraction, summarization, conversation compaction, and subagent duty underneath a larger planner model such as Claude Opus 5.5 or Claude Sonnet 5.5. Anthropic explicitly notes that the larger models remain the better fit for complex agentic coding, while Haiku 5.5 targets narrowly scoped tasks that were previously cost-prohibitive at scale.

Against its own predecessor, Haiku 4.5, the changes are concrete. The context window grows from 200K to 1M tokens and maximum output from 64K to 128K, with up to 300K output tokens available on the Message Batches API behind a beta header. It is also the first Haiku to expose effort controls, with adaptive thinking on by default and configurable effort levels; thinking tokens count toward the output limit, so existing token caps may need revisiting.

One migration detail matters: Haiku 5.5 uses the newer tokenizer shared by Claude 4.7 and later models, so identical input text counts roughly 30% more tokens than on Haiku 4.5.

On Venice, the model is offered with reasoning, vision, code-optimized behaviour, function calling, and web search enabled, suited to pipelines that run many short requests rather than a few long ones.