Claude 4.5 Haiku (Non-reasoning)

Released by Anthropic on October 15, 2025, Claude 4.5 Haiku is the high-speed, cost-efficient entry in the Claude 4.5 model family. It is designed to provide near-frontier performance, specifically matching the coding capabilities of the earlier Claude Sonnet 4 model, while operating at significantly higher speeds and a lower price point. The non-reasoning designation typically refers to its operation in a standard low-latency mode without the "Extended Thinking" feature enabled, prioritizing immediate throughput for real-time applications.

The model is optimized for high-volume, latency-sensitive tasks such as customer service automation, real-time data streaming, and agentic workflows. It excels in computer use and sub-agent orchestration, where it can be used to execute discrete subtasks planned by larger models like Claude 4.5 Sonnet. On the SWE-bench Verified coding benchmark, the model achieves a score of approximately 73.3%, placing it among the most capable lightweight models for software development.

Claude 4.5 Haiku features a 200,000-token context window and a maximum output limit of 64,000 tokens. Its knowledge cutoff is dated to July 2025. The architecture is built to be context-aware, allowing the model to track its own token usage and manage persistence across long conversations to reduce agentic "laziness."

From a safety perspective, the model is classified under Anthropic's AI Safety Level 2 (ASL-2). Internal evaluations indicate that it demonstrates lower rates of misaligned behaviors than some larger frontier models, which Anthropic attributes to specialized alignment training and its optimized scale.