Anthropic Launches Claude Haiku 4.5 – Its Fastest, Most Economical, and Safest AI Model Ever

AI company Anthropic has just officially introduced Claude Haiku 4.5, the latest member of the Claude 4 model line. The company describes this as the fastest, most cost-effective and safest version in its entire product portfolio.

 

According to the announcement, Claude Haiku 4.5 costs only $1 for every 1 million tokens input and $5 for every 1 million tokens output, greatly reducing operating costs for businesses and developers.

Twice the speed, one-third the cost of the Sonnet 4.5

Anthropic says Haiku 4.5 achieves nearly the same programming and reasoning capabilities as the Claude Sonnet 4.5 – a larger model – but is twice as fast and up to three times cheaper.

Because of this advantage, Haiku 4.5 is especially suitable for real-time applications, such as chat assistants (chatbots), customer care support tools, or AI supporting pair-programming.

 

On the SWE-bench Verified performance benchmark – a dataset specifically used to test the programming capabilities of AI models – Claude Haiku 4.5 scored 73.3%, just slightly lower than Sonnet 4.5 (77.2%).

 

Flexible model coordination and efficient multitasking

Anthropic also says that Haiku 4.5 has significantly improved sub-agent orchestration. For example, a single Claude Sonnet 4.5 instance can take on the role of overall project manager, while smaller tasks are split across multiple Haiku 4.5 instances for faster processing.

This approach makes AI more flexible, efficient, and resource-efficient, especially in complex processes.

Improved safety – AI Safety Level 2 certified

On the safety front, Anthropic claims that Claude Haiku 4.5 exhibits significantly less erratic behavior than previous versions. Results from automated tests show that the model meets AI Safety Level 2 (ASL-2) safety standards – less restrictive than ASL-3, which is applied to larger models like Sonnet 4.5 and Opus 4.1.

Anthropic says this makes Haiku 4.5 its safest model yet, particularly in preventing AI misuse to create sensitive or dangerous content, such as chemical, biological, or nuclear data.

 

Developer-ready across multiple platforms

Developers can now access Claude Haiku 4.5 through the Claude API, or the Amazon Bedrock and Google Cloud Vertex AI platforms.

Anthropic also ensures backward compatibility for users deploying earlier versions of Haiku 3.5 or Sonnet 4.