Haiku 4.5
The fourth-generation small AI language model in Anthropic's Haiku series, serving as the predecessor to Claude Haiku 3.5.
Core Facts
Timeline (last 90 days)
Haiku 5.5 achieved a score of 1620 on the relevant coding benchmark, far higher than Haiku 4.5's score of 735.
On the Artificial Analysis Intelligence Index, Haiku 4.5 scored 17 points, while Haiku 5.5 jumped to 43 points, second only to Kimi K3's 44 points.
On the GDPVal benchmark, Haiku 5.5 scored 1620 and Haiku 4.5 scored 735.
The Claude apps gateway's backend requests now use Haiku 4.5 instead of the session model to reduce overhead.
In Databricks' OfficeQA Pro V1 evaluation, Claude Haiku 5.5 achieved approximately 15% quality improvement over the previous generation Haiku 4.5, while costing only a fraction of the latter.
Haiku 5.5's input price has dropped to $0.10 per million tokens
Compared to Claude Haiku 4.5, Haiku 5.5 reduces operating costs by an average of approximately 75%.
In a Wikipedia Speedrun demonstration, Jev completed five hops in just half a second, while Tara 5.6, Haiku 4.5, and Sonnet 5 took approximately 4 seconds, 5 seconds, and 5 seconds respectively.
论文在三大厂商验证有效：Anthropic用Haiku 4.5解码Opus 4.8，OpenAI用GPT-5 Mini解码GPT-5，Google用Gemini解码其Flash版本，且解码后推理链token数量与API报告完全吻合
All Facts (9)
Compared to Claude Haiku 4.5, Haiku 5.5 reduces operating costs by an average of approximately 75%.
90%VerifiedHaiku 5.5's input price has dropped to $0.10 per million tokens
70%UnverifiedHaiku 5.5 achieved a score of 1620 on the relevant coding benchmark, far higher than Haiku 4.5's score of 735.
50%UnverifiedOn the GDPVal benchmark, Haiku 5.5 scored 1620 and Haiku 4.5 scored 735.
50%UnverifiedOn the Artificial Analysis Intelligence Index, Haiku 4.5 scored 17 points, while Haiku 5.5 jumped to 43 points, second only to Kimi K3's 44 points.
50%UnverifiedThe Claude apps gateway's backend requests now use Haiku 4.5 instead of the session model to reduce overhead.
50%UnverifiedIn Databricks' OfficeQA Pro V1 evaluation, Claude Haiku 5.5 achieved approximately 15% quality improvement over the previous generation Haiku 4.5, while costing only a fraction of the latter.
50%UnverifiedIn a Wikipedia Speedrun demonstration, Jev completed five hops in just half a second, while Tara 5.6, Haiku 4.5, and Sonnet 5 took approximately 4 seconds, 5 seconds, and 5 seconds respectively.
50%Unverified论文在三大厂商验证有效：Anthropic用Haiku 4.5解码Opus 4.8，OpenAI用GPT-5 Mini解码GPT-5，Google用Gemini解码其Flash版本，且解码后推理链token数量与API报告完全吻合
50%