کلوڈ ہائیکو 4.5

The fastest Claude, delivering near-frontier coding quality with very low latency and cost. It is built for real-time work -- chat assistants, customer service agents, pair programming -- and surpasses Claude Sonnet 4 at some tasks, including computer use, at far lower cost. It has no effort control: thinking is the older extended-thinking mode, off unless a budget is requested.

Anthropic's pitch for Haiku 4.5 was a statement about how fast the frontier moves: five months earlier Claude Sonnet 4 was state of the art, and Haiku 4.5 delivers similar coding performance at a third of the cost and more than twice the speed. It was released two weeks after Sonnet 4.5, which Anthropic was careful to say remained its frontier model; Haiku 4.5 exists as the option for near-frontier work where cost and latency decide the design.

The benchmark table backs the framing rather than the headline. Anthropic reports 73.3% on SWE-bench Verified against 72.7% for Sonnet 4, 41.0% on Terminal-Bench against 36.4%, and 50.7% on OSWorld against 42.2% - computer use is where the gap over Sonnet 4 is widest. Against its own larger sibling Sonnet 4.5 it lands lower across the board, at 73.0% on GPQA Diamond against 83.4% and 83.0% on MMMLU against 89.1%, which is the trade the tier is for.

The structural idea Anthropic put forward with it is orchestration rather than substitution: a larger model breaks a complex problem into a multi-step plan, then directs a team of Haiku 4.5 instances working subtasks in parallel. That pattern is why speed matters more than peak score here, and why the lab pointed at Claude Code and browser-based work as the places the difference is felt.

Thinking on this model is the older extended-thinking mode rather than an effort dial, and the lab's own methodology shows what that means in practice. Its published Terminal-Bench figure averages eleven runs, six of them with no thinking at all scoring 40.21% and five with a 32K thinking budget scoring 41.75%; the SWE-bench Verified number was produced with a 128K budget. Thinking is off unless a call requests a budget, and on this model the budget is the only control.

Safety is where Haiku 4.5 stood out most at release. Anthropic reports it as substantially more aligned than Haiku 3.5 and, on the automated alignment assessment, showing a statistically significantly lower overall rate of misaligned behaviour than both Sonnet 4.5 and Opus 4.1 - by that one metric, the safest model the lab had shipped. Its testing also found only limited risk around chemical, biological, radiological and nuclear weapons, so it was released under the ASL-2 standard rather than the stricter ASL-3 applied to Sonnet 4.5 and Opus 4.1.