Today marks the release of Anthropic’s latest compact model, Claude Haiku 4.5. Priced at a mere one-third of Claude Sonnet 4, it delivers over twice the processing speed.

Claude Haiku 4.5 also significantly boosts response times in coding scenarios, even outperforming Claude Sonnet 4 in select computational tasks. These enhancements render applications like Claude for Chrome more practical than ever before.

Although its overall performance may not quite reach the level of Sonnet 4.5, Claude Haiku 4.5 offers a compelling new choice for users seeking near-cutting-edge capabilities paired with cost-efficiency. Currently, it is priced at $0.2 per million input and output tokens.

This model also paves the way for innovative model combinations. For instance, Sonnet 4.5 can dissect a complex problem into a multi-step plan and then orchestrate a team of multiple Haiku 4.5 models to tackle subtasks in parallel.

In the realm of agent coding, Haiku 4.5 has outperformed both Sonnet 4 and ChatGPT 5 (high), achieving an impressive accuracy rate of 73.3%.

During mathematical ability assessments, while Haiku 4.5’s success rate doesn’t quite match that of ChatGPT 5 and Sonnet 4, it still attains an accuracy rate of 96.3% when utilizing tools.

When it comes to visual reasoning, Haiku 4.5’s capabilities are comparatively weaker, ranking the lowest among several evaluated models.

Nevertheless, on the whole, users are highly satisfied with Haiku 4.5’s performance and have given it glowing reviews.

In terms of safety, Claude Haiku 4.5 demonstrates a low occurrence of concerning behaviors and exhibits notably improved consistency compared to Claude Haiku 3.5.

Automated consistency evaluations reveal that Claude Haiku 4.5’s overall incidence of biased behaviors is also substantially lower than that of Claude Sonnet 4.5 and Claude Opus 4.1. Consequently, based on this metric, Claude Haiku 4.5 stands as the safest model Anthropic has developed to date.

Safety tests further indicate that Claude Haiku 4.5 poses minimal risks concerning the production of chemical, biological, radiological, and nuclear (CBRN) weapons.

Claude Haiku 4.5 is now accessible on Claude Code and Anthropic’s applications. Developers can leverage Claude Haiku 4.5 through Anthropic’s API, Amazon Bedrock, and Google Cloud’s Vertex AI.

References: https://www.anthropic.com/news/claude-haiku-4-5