Frontier models should spend their time thinking.

The best AI models can now reason through hard problems. Watch one work through a coding task, though, and most of its time goes elsewhere: reading files, searching code, making routine edits, running tests.

That work matters, but it doesn’t need the most capable model there is. Running it there is slow and costly, and it pulls the model away from the decisions only it can make.

We’re building specialists for that work. A specialist takes a defined job from the frontier model, does it much faster, and reports back. The frontier model stays in charge of the plan and the judgement.

Soru-1 is our first, for coding. It works as a sub-agent inside the tools you already use, and we’ll publish how it performs when it launches.

We think this is how intelligence should be put to work: the right model on each part of the job.