Claude Haiku 4.5

Overview. Claude Haiku 4.5 is Anthropic's fastest model in the Claude 4 generation, publicly released on 2025-10-15. It's built for developers who need quick, high-throughput Inference without giving up tool-use and coding competence, and it accepts text, image, and file input with a 200K-token context window. Anthropic doesn't release model weights, so this is served as a closed, hosted model rather than something you can self-host.

Key strengths. On SWE-bench Verified, Haiku 4.5 scores 73.3%, and on BFCL-v4 it scores 68.7%, putting coding and agentic tool-calling as its clearest strengths. It supports Anthropic's structured tool-use schema, including flag-driven parallel tool calls, and optional extended thinking for problems that benefit from a visible reasoning step. Anthropic also reports strong internal safety-evaluation results for this model, which supports its use in sensitivity-conscious deployments.

When to use it. Reach for Haiku 4.5 when you need low-latency, real-time responses for coding assistants, agentic workflows, or tool-calling pipelines rather than deep multi-step research tasks. It's a solid fit for document- and image-aware assistants that read PDFs or screenshots and need to act on them quickly. Its optional extended-thinking mode makes it flexible enough for moderately complex tasks without paying the latency cost of always-on reasoning.

Limitations. Haiku 4.5 is positioned by Anthropic as the speed-optimized member of the Claude 4 family, trading some raw capability for lower latency compared to Sonnet- or Opus-class models, so it's not the right choice for the hardest reasoning or research tasks. It has no audio input or output, and its training data runs only through July 2025. Structured JSON output and parallel tool calls are opt-in configuration choices rather than always-on defaults.

Sovereignty note. Claude Haiku 4.5 is built by Anthropic, a US-based creator, which means the model carries CLOUD Act exposure rather than EU-Sovereign status. Treat data residency and cross-border access considerations accordingly when routing sensitive workloads to this model.