Claude 4.5 Haiku (Reasoning) is the compact, speed-optimized model in the Claude 4.5 family developed by Anthropic. Released as a successor to previous lightweight iterations, it is designed for high-throughput applications that require near-frontier intelligence at a lower cost and latency. It is the first model in the Haiku class to incorporate Extended Thinking capabilities, allowing it to toggle between rapid execution for routine prompts and deliberate internal reasoning for complex challenges.
Capabilities and Performance
The model features a hybrid reasoning architecture that enables it to perform internal chain-of-thought processing before generating a final response. This approach allows it to handle multi-step coding tasks, mathematical deductions, and scientific problem-solving with higher accuracy than previous small-scale models. In benchmark evaluations, the model achieved a 73.3% score on SWE-bench Verified, matching the coding performance of larger models like the original Claude 4 Sonnet while operating at approximately twice the speed.
Architecture and Context
Claude 4.5 Haiku supports a 200,000-token context window and is multimodal, capable of processing both text and image inputs. It utilizes a training intervention known as "explicit context awareness," which provides the model with precise information about its context usage. This feature is intended to minimize agentic "laziness" by ensuring the model continues reasoning persistently through long-horizon tasks and wraps up responses gracefully as it nears the context limit. Its knowledge cutoff is February 2025.