Anthropic’s Claude Haiku 4.5, a newly released low-cost large language model (LLM), represents the company’s answer to growing market demand for cheap artificial intelligence (AI) solutions.
The San Francisco-based company is pricing the model at $1 per million input tokens and $5 per million output tokens, a fraction of what it charges for Claude Sonnet 4.5, its flagship model, which costs roughly three times as much.
Rather than sacrifice performance, the company engineered Haiku 4.5 as a hybrid reasoning model capable of adjusting computational resources based on task complexity. By default, the algorithm operates in a lightweight mode that requires minimal hardware. Users seeking more sophisticated analysis can activate an extended thinking feature that triggers more intensive processing for complex queries.
Anthropic drew from multiple sources to train the model, including publicly available web content, third-party data providers, and internal datasets. The company included information from customers who consented to have their data used for AI training purposes. The training dataset was scrubbed of duplicates to optimize efficiency.
The model can handle substantial volumes of information, accepting multimodal prompts containing up to 200,000 tokens, which is enough to process lengthy business documents and other large files. Haiku 4.5 can generate responses up to 64,000 tokens in length.
When subjected to eight standard industry benchmarks, Haiku 4.5 came within 10% of Sonnet 4.5’s performance across most tests. More notably, it surpassed the company’s previous flagship model, Sonnet 4, on three benchmarks focused on coding and high school mathematics problems.
Beyond its competitive pricing, Anthropic is marketing Haiku 4.5 as the company’s safest model to date. The algorithm also operates more than twice as fast as Sonnet 4, making it particularly attractive for applications where response speed matters, such as customer support systems and AI agent workflows.
The latter application has caught Anthropic’s attention. According to the company, developers can build agents using Sonnet 4.5 but delegate routine tasks to Haiku 4.5 sub-agents, potentially slashing inference costs while automating multi-step processes like coding projects and market research.
Anthropic is making Haiku 4.5 available through its application programming interfaces and Claude.ai chatbot, as well as Claude Code, the company’s programming assistant that has emerged as a significant revenue engine since its May launch. Reuters reported this week that Claude Code is approaching $1 billion in annual recurring revenue.
The strategy mirrors moves by competitor OpenAI, which offers scaled-down versions of its flagship model, including GPT-5 Mini and GPT-5 Nano, both priced lower than their full-featured counterpart. Both companies have also enabled developers to cache frequently used prompts, reducing costs by eliminating redundant computations.
