Claude「最香」模型发布，速度翻倍价格大砍，编程能力直逼 GPT-5

Core Insights - Anthropic has released Claude Haiku 4.5, which offers improved performance, faster speed, and lower pricing compared to its predecessors [1][2][9] - Haiku 4.5 achieves a score of 73% on the SWE-bench Verified test set, comparable to Claude Sonnet 4 and OpenAI's GPT-5 [4][8] - The model is particularly effective for real-time, low-latency tasks, enhancing user experience in applications like chat assistants and coding [6][10] Performance Comparison - Haiku 4.5's pricing is set at $1 per million input tokens and $5 per million output tokens, making it significantly cheaper than Sonnet 4.5, which is approximately three times more expensive [9][11] - In specific tasks, Haiku 4.5 outperforms Sonnet 4, showcasing its capability in handling complex problems [5][7] - Compared to GPT-5 mini and Google’s Gemini 2.5 Flash, Haiku 4.5 is about four times more expensive, but offers better performance in scenarios requiring tool usage and autonomy [9][10] Safety and Alignment - Haiku 4.5 has undergone extensive safety and alignment testing, showing a lower incidence of undesirable behavior compared to its predecessor, Haiku 3.5 [7] - The model's overall deviation behavior is less than that of Sonnet 4.5 and Opus 4.1, indicating it is the safest model in Anthropic's lineup [7] Future Developments - Anthropic plans to expand its capabilities beyond programming assistants, potentially acquiring technologies for automated code testing and software design tools [13][15] - The company is also considering small acquisitions under $500 million to enhance its offerings in specific industries like financial services and healthcare [15]