Google hands the cheaper model part of the coding job
Google’s two-model system finished a coding benchmark faster while using less quota. The savings depend on which steps can change hands.
Google’s Antigravity Orchestrator can switch between Gemini 4 Carbon and Hydrogen during a single session. Google reports that pairing them on Terminal-Bench 4.0 delivered a success rate close to Carbon alone, finished 19% faster on average and used 24% less quota.1
The unit that matters
A software job can include planning, editing and checking a result. Routing lets a system spend more on a difficult decision and less on routine steps. A wrong handoff can erase those savings if the stronger model has to repeat the work.
Google reports the routing figures for a particular coding benchmark and measures quota inside its product. Translating that into a company’s dollar savings would require its own mix of tasks, error rates and review costs.1