The interesting thing in Cua’s CUA-S1 release isn’t the model — it’s the architectural bet underneath it. The team asked how many computer-use steps actually need a general-purpose agent burning tokens on gpt-6-astra or claude-opus-5, and how many are just local decisions: put this value in that box, check this one, skip that field. Their answer is a 706k-parameter option-scorer that doesn’t generate tokens at all. It takes the current context plus a fixed set of choices and returns a probability per choice — something your code can inspect, trust, and act on.
That framing lands for anyone running agents in production. Most of my orchestration pain isn’t the model failing to reason; it’s the general agent loop being invoked for decisions that were never ambiguous, paying full latency and cost to conclude the obvious. CUA-S1-FORMS scores a whole form in 7-9 ms locally against ~260-280 ms per hosted call, and on their scoped form task it beats the general baseline on exactly the boring cases — 100% vs 74% on “leave the already-filled field alone.” Scoped specialization on a narrow decision space is doing the work here, not scale.
The honest caveats are all there: it’s forms-only, trained on synthetic data, ignores screenshots, and won’t predict new text values. But the direction is the part worth stealing — a general agent that hands well-understood decisions to cheap, checkable specialists and keeps its expensive reasoning for genuine novelty. That’s a routing problem, not a model problem.
The HN discussion is worth reading for where people think the specialist/generalist boundary actually sits. Where in your agent stack are you paying for system-2 reasoning on system-1 decisions?