Anthropic Releases Claude Haiku 4.5 as a Fast and Cost-Effective AI Model

Anthropic, on Wednesday, introduced the Claude Haiku 4.5 artificial intelligence (AI) model. It is the smallest model in the Claude 4.5 series, and the second to be launched after the Claude Sonnet 4.5. Unlike the frontier model, its focus is on low-latency tasks, with the San Francisco-based AI firm highlighting that it is twice as fast as the Sonnet 4, while delivering similar levels of coding performance. Just like Sonnet 4.5, the large language model (LLM) is available to users on the free tier.

Anthropic Debuts Claude Haiku 4.5 AI Model

In a newsroom post, the AI firm announced and detailed its latest small AI model, highlighting benefits such as faster speed and cheaper application programming interface (API) costs as the main benefits. Unlike Sonnet 4.5, it is not a frontier model, but is geared towards real-time use cases due to its low latency.

Anthropic claims that Claude Haiku 4.5 surpasses the last generation's frontier model in tasks such as Computer Use, which makes products such as Claude for Chrome (currently in beta) more useful. The company added that “Users who rely on AI for real-time, low-latency tasks like chat assistants, customer service agents, or pair programming will appreciate Haiku 4.5.”

Coming to benchmarks, the AI firm shared the scores achieved by the model during internal evaluation. As per the shared results, it outperforms Google's Gemini 2.5 Pro in the SWE-bench Verified benchmark for coding performance, and Terminal-bench for agentic terminal coding. It also beats both Gemini 2.5 Pro and GPT-5 in Tau-2 for agentic tool use and OSWorld for computer use benchmarks. However, its performance in MMMU for visual reasoning, GPQA Diamond for reasoning, and AIME 2025 for mathematics is below the frontier models.

On safety evaluation, Anthropic said it is more aligned than its predecessor, Claude Haiku 3.5, based on its automated alignment assessment. The company also called it “our safest model yet” by highlighting that it showed a lower overall rate of misaligned behaviour than both Claude Sonnet 4.5 and Claude Opus 4.1. It has been released under the AI Safety Level 2 (ASL-2) standard for limited chemical, biological, radiological, and nuclear (CBRN) risks.