Claude Haiku 4.5: fast, cheap model, free on Claude.ai

What's new

On October 15, 2025, Anthropic launched Claude Haiku 4.5. Haiku is the company's small, fast model series, and with version 4.5 it has reportedly become so capable that it comes close to the large models.

The model can read up to 200,000 tokens at a time (tokens are the chunks of text an AI works in) and can write up to 64,000 tokens in a single response. It supports extended thinking, where the model spends more time thinking a task through. Its knowledge extends to February 2025.

Haiku 4.5 is free for all users on Claude.ai. It is also available in Claude Code, Anthropic's apps, the API, Amazon Bedrock and Google Cloud Vertex AI. At GitHub Copilot, it is in public preview for Pro, Pro+, Business and Enterprise subscribers. It can be selected in VS Code, on github.com and in GitHub Mobile, and the rollout is happening gradually.

What's smart

Speed is the big draw. Anthropic says Haiku 4.5 is more than twice as fast as Sonnet 4. Customer testimonials from the launch even mention four to five times the speed of Sonnet 4.5, but those are statements, not an independent measurement.

According to Maginative, the model also performs well on computer use, meaning when the AI operates buttons, menus and browsers on the screen the way a human would.

Anthropic also points to a collaboration between the models: Sonnet 4.5, which remains the company's leading model, can plan a large task and direct several Haiku 4.5 models, each solving its own part at the same time.

On safety, the model is classified ASL-2 and, according to Anthropic's own evaluation, showed lower rates of misaligned behavior than both Sonnet 4.5 and Opus 4.1.

Cheaper or better?

It is clearly cheaper. Through the API, Haiku 4.5 costs 1 dollar per million input tokens and 5 dollars per million output tokens. Sonnet 4 and 4.5 cost 3 and 15 dollars, three times as much. For comparison, GPT-5 costs 1.25 dollars and Opus 4.1 costs 15 dollars per million input tokens. A monthly price for regular users is not relevant on Claude.ai, where the model is free.

On the SWE-bench Verified coding benchmark, Haiku 4.5 scored 73.3 percent. Sonnet 4 got 72.7 percent and Google's Gemini 2.5 Pro 67.2 percent, according to Maginative's account. SD Times writes that Haiku 4.5 beats Sonnet 4 but trails Sonnet 4.5. All performance and speed figures ultimately come from Anthropic and have not been independently measured.

What it's good for

Anthropic points to tasks where the answer needs to come quickly: chat assistants, customer service and pair programming, where the AI sits alongside the developer and helps with the code as they go.

For regular users, this means a fast, free assistant on Claude.ai. For businesses, it can mean noticeably lower bills if they currently use a more expensive model for tasks that a smaller one can handle.

In GitHub Copilot, administrators for Business and Enterprise must enable a policy for the model before employees can use it. The sources do not say whether the model has special terms in Denmark or the EU.