Anthropic Unveils New, More Efficient Version of Haiku Model

On Wednesday, Anthropic unveiled Claude Haiku 4.5, the latest update to its smallest model — now promising performance on par with Sonnet 4, but “at one-third the cost and more than twice the speed,” according to the company’s blog post.

The company backed up its claims with a series of new benchmark results. In internal testing, Haiku 4.5 scored 73% on SWE-Bench verified and 41% on the Terminal-Bench, a command-line-focused test — slightly behind Sonnet 4.5, but comparable to Sonnet 4, GPT-5, and Gemini 2.5 across multiple categories. Similar parity was seen in evaluations of tool use, computer use, and visual reasoning.

Anthropic said the upgraded Haiku model is now available across all free Anthropic plans, emphasizing its potential value for AI products that prioritize low-cost, high-speed performance. Thanks to its lightweight design, Haiku 4.5 can be deployed easily in parallel or paired with more powerful models, enabling scalable and efficient AI systems.

In a statement to the press, Anthropic CPO Mike Krieger described Haiku 4.5 as a breakthrough for real-world deployments. “It’s opening up entirely new categories of what’s possible with AI in production environments — with Sonnet managing complex planning while Haiku-powered sub-agents execute rapidly,” he said. “We’re providing a complete agent toolbox that balances intelligence, speed, and cost for every stage of a task.”

The company expects early adoption in software development tools, where latency and responsiveness are critical. Zencoder CEO Andrew Filev praised the release, calling Haiku 4.5 “a catalyst for an entirely new range of use cases.”

The launch of Haiku 4.5 continues Anthropic’s rapid product rollout cycle — arriving just two weeks after Sonnet 4.5 and two months after Opus 4.1, both of which were widely regarded as state-of-the-art models. The previous Haiku version debuted in October 2024, marking a significant leap forward in less than a year.