A coding agent doesn’t need its strongest model for every step

In Google’s description of Orchestrator, Carbon or Argon starts by looking at the code and making a plan. Hydrogen handles edits and tests. If it runs into something the plan didn’t account for, the planner can take over again without starting a new session. Which planner you can choose depends on your plan: Google’s model docs list Carbon for Ultra and Argon for Ultra and eligible Pro users.

That’s different from picking a model once at the start. The first prompt tells the agent what you want. It might not tell the agent why the job is hard. Sometimes the failed test is the first clue.

Google reported nearly the same task success as Carbon alone in its benchmark, with runs finishing 19% faster on average and using 24% less quota. Those are Google’s results for that benchmark. They don’t say how often the handoff works on your codebase, or whether a faster first pass saves you time after review.

A good first test is a small change with a known complication. Use a branch where you already know which older behavior has to survive. Give the same request to Orchestrator and to the planner model on its own. Check whether each notices the complication, preserves that behavior, and runs the test that would catch a regression. Then compare how much time you spend reading and fixing the result.

If both get it right, the faster run is useful. If only one notices that the old config still has to load, that’s the difference to pay attention to. The point isn’t to use the stronger model less. It’s to use it when the job calls for it.