Who plans, who executes

Google released Gemini 4 Hydrogen yesterday, a fast and cheaper model for the hands-on work of agents: writing code, running tests, checking whether a change is correct. Google leaves the heavy reasoning to its larger models, Carbon and Argon.

Alongside it comes Orchestrator in Antigravity, which switches between models within a single session. According to Google, on Terminal-Bench 4.0 this achieves almost the same success rate as Carbon on its own, with 24 percent lower usage on average. For parallel work, Carbon can coordinate while Hydrogen directs builders and reviewers, each in its own workspace. The price: 1.50 dollars per million input tokens.

All figures come from Google itself. Independent tests are still to come.

It looks more and more like a team: one that plans and reviews, many that carry out the work. That raises a question we already know from people. If one model plans and another executes, who answers for the result?