Today, Anthropic announced the official launch of Claude Haiku 4.5, a small-scale AI model. This model delivers powerful performance approaching that of cutting-edge models at a very low cost, making it an ideal choice for low-latency tasks such as real-time chat assistants and online customer service. As the smallest member of the Claude family, Haiku 4.5 leverages a "distillation" technique to achieve encoding capabilities comparable to larger models while significantly reducing operating costs.
In terms of performance, Haiku 4.5's encoding capabilities are roughly equivalent to the medium-sized Sonnet 4 model, at only one-third the cost and over twice the processing speed. According to data from the authoritative benchmark SWE-bench Verified, Haiku 4.5 scored 73.3%, slightly higher than Sonnet 4's 72.7%. Even more impressively, in some tasks that simulate human-like computer operation, its performance even surpassed Sonnet 4 and approached the performance of OpenAI's GPT-5. However, officials caution that these data may be filtered and should be interpreted with caution.
Pricing is a core competitive advantage of Haiku 4.5. For developers, its API call price is $1 per million input tokens and $5 per million output tokens, significantly lower than Sonnet 4.5's $3/$15 and Opus 4.1's $15/$75. Anthropic has also innovatively designed a multi-model collaborative workflow, where Sonnet 4.5 decomposes complex tasks and then schedules them for parallel execution across multiple Haiku 4.5 instances, significantly improving efficiency. This architecture opens up new possibilities for advanced scenarios such as AI agent-based encoding, further solidifying the Haiku family's position as a cost-effective alternative.
