Anthropic introduced Claude Haiku 4.5 on Wednesday, unveiling the latest iteration of its most compact model. According to the company’s blog, this version delivers performance comparable to Sonnet 4, but at just a third of the price and at more than double the speed.
To support these claims, Anthropic shared various new benchmark outcomes. Their internal evaluations show Haiku achieving a 73% score on SWE-Bench verified and 41% on Terminal-Bench, which focuses on command-line tasks—results that trail Sonnet 4.5 but match those of Sonnet 4, GPT-5, and Gemini 2.5. Comparable performance was observed in tests involving tool usage, computer tasks, and visual reasoning.
Haiku’s latest version is now instantly accessible to all users on Anthropic’s free plans. The company expects it to be especially attractive for free AI product offerings, as it delivers robust features while reducing server demand. Its lightweight design also allows for the simultaneous deployment of several Haiku agents, or for pairing with more advanced models.
Anthropic’s Chief Product Officer, Mike Krieger, told the press that Haiku enables new deployment approaches in production settings. “This unlocks whole new possibilities for AI in real-world applications—Sonnet can manage intricate planning, while Haiku-driven sub-agents operate rapidly,” Krieger explained. “We’re providing a comprehensive agent toolkit, letting each model offer the optimal mix of intelligence, efficiency, and cost for specific tasks.”
The most immediate impact is expected in software development tools, where Claude Code is already widely adopted and low latency is crucial. Zencoder CEO Andrew Filev, quoted by Anthropic, said the updated Haiku “opens up a brand new range of use cases.”
The release of Haiku 4.5 follows several major product launches by Anthropic, arriving just two weeks after Sonnet 4.5 and two months after Opus 4.1—both of which were considered cutting-edge at their debuts. The prior Haiku version was launched in October 2024.