Claude Haiku 4.5: Fast AI at a Lower Cost

Anthropic now offers Claude Haiku 4.5, an AI model designed to approach the performance of larger models without costing or taking as much time. According to the company, it reaches a level of programming performance similar to Claude Sonnet 4, costs one-third as much and responds more than twice as fast.

The priority here is not solving a single task with maximum depth, but working without pauses. That matters when AI maintains a conversation, handles customer requests, helps with programming or controls applications across multiple steps.

What Haiku 4.5 offers

Claude Haiku 4.5 is designed for tasks that need fast, repeated responses, such as:

• Chat assistants that must respond in real time.
• Customer service agents.
• Assisted programming and rapid prototyping.
• Agents that use a computer and complete actions across different applications.
• Teams of multiple agents that divide complex work into subtasks.

Anthropic says Haiku 4.5 even outperforms Claude Sonnet 4 on some computer-use tasks. It also says the model achieves 90% of Sonnet 4.5's performance in an agentic programming evaluation conducted by Augment, although that result comes from a specific test and does not represent every possible use case.

The idea is to combine models. Claude Sonnet 4.5 can analyze a problem and break it into steps, while multiple instances of Haiku 4.5 execute those tasks in parallel. For you, that could mean applications that respond sooner and development processes that cost less.

Price and availability

Haiku 4.5 is available starting today in Claude, Claude Code and Anthropic's applications. Developers can access it through the API, Amazon Bedrock and Google Cloud Vertex AI.

The price is $1 per million input tokens and $5 per million output tokens. Tokens are the small units of text the model receives and generates. In practice, the final cost depends on how much text your application processes and how many times it calls the model.

Anthropic also presents it as a direct replacement for Haiku 3.5 and Sonnet 4 in some workflows. That lets you try it without completely redesigning an application that already uses those models.

It does not replace the most powerful model

Claude Sonnet 4.5, introduced two weeks earlier, remains Anthropic's reference model for programming and complex tasks. Haiku 4.5 occupies a different position: it offers performance close to that of advanced models when speed and price matter more than maximum capability.

The difference is similar to choosing between a specialist who spends more time analyzing each case and a fast team that resolves many tasks at once. For an application processing thousands of requests, that difference may matter more than achieving the latest possible improvement in a test.

What Anthropic says about safety

Anthropic says Haiku 4.5 showed fewer problematic behaviors than Haiku 3.5 in its evaluations, as well as a lower rate of misaligned behavior than Sonnet 4.5 and Opus 4.1 in an automated test of its own. These are company results, not an independent certification.

The model has been classified at ASL-2, because Anthropic considers its risks related to the production of chemical, biological, radiological and nuclear weapons to be limited. Sonnet 4.5 and Opus 4.1 have a more restrictive classification, ASL-3.

The arrival of Haiku 4.5 shows where applied AI is heading: not only toward more capable models, but toward models that are good enough to run many times, with little waiting and at an affordable cost. The important question will be how much of that advantage holds up in real applications, outside controlled tests and the figures provided by the manufacturer.