Text, code, images: every AI task in one place

The best-answer principle doesn't stop at chat. Anywhere multiple models can attempt the same task, comparing their outputs and keeping the winner beats trusting any single one — and that applies to far more than text.

Code that actually runs

For programming tasks, candidate answers are judged on correctness first: does the code do what was asked, does it handle edge cases, is it idiomatic? Explanatory quality is scored separately, so you get code that works and a walkthrough you can actually follow.

Images and research

Image generation runs across multiple image models, and the judge scores prompt fidelity and visual quality. Research tasks weight sourcing and factual grounding heavily, so confident-sounding nonsense gets filtered out rather than surfaced.

One subscription, one interface, every kind of AI work — with the model competition happening silently behind the scenes on every request. That's the whole idea.