Computer use is now cheap. The bottleneck shifts from capability to cost gating.

Google added built-in computer use to Gemini 3.5 Flash. That means the same model you can run for a fraction of a cent per 1k tokens now ships with a native tool to see a screen, click, type, and scroll. No orchestrator needed.

I’ve been running agents that interact with UIs for a while now. The pattern I landed on was an architecture where the default question every tick is whether the model needs to be called at all (An agent built around not calling the LLM). That choice came directly from token costs. Screen captures chew through context; every screenshot you take adds thousands of tokens of vision input, and an agent that takes 40 screenshots to book a flight burns money on a problem that isn’t reasoning—it’s a loop that didn’t know when to stop.

Flash changes the math. The model is fast enough and cheap enough that the same 40-screenshot task might cost less than my morning coffee. What I’m less convinced about is whether that means we should throw away gating entirely. Cheap calls still add up across long-horizon tasks, and cheap calls don’t fix the brittleness—an agent that gets stuck retrying a grayed-out button still burns budget for no progress.

The built-in tool is a step forward because it removes the need to stitch screen-capture libraries and prompt engineering. But what I want to see next isn’t a faster model attached to pixels. It’s tool-level controls: a way to say “take at most N screenshots before pausing and asking me” or a native backoff when the screen hasn’t changed. Without those, cheap computer use gets you more ways to waste tokens, not necessarily more things done.

I’ll test it on a few real flows I already have—form-filling, dunning letter checks, the kind of stuff where spelling errors cost money. If the Flash tool lets me replace the screen-rate gating I built myself, I’ll swap. If not, I’ll keep the gating and just drop a cheaper model behind it.