Google announced Gemini 4 Neon on Nov. 18, positioning it for complex software engineering, analytical work and production-grade workflows. The release gives it a one-million-token context window and API list prices of $3 per million input tokens and $15 per million output tokens. Google places Neon between Carbon for harder, long-horizon work and Hydrogen for fast everyday tasks, and says Neon can run alone or act as an Antigravity orchestrator directing parallel Hydrogen subagents. Those are product and vendor claims, not a measure of cost per accepted code change. S1

The routing claim has a specific scope

The mechanism predates Neon. In a Nov. 12 Antigravity post, Google described Orchestrator as a dynamic handoff between Carbon and Hydrogen within a single task trajectory: Carbon can plan, Hydrogen can implement and test, and Carbon can step back in when the worker hits friction. Google reported near-Carbon task success on Terminal-Bench 4.1, using 24% less quota and finishing 19% faster on average. S2 Those numbers refer to that Carbon/Hydrogen Orchestrator setup—not to Neon, and not to Terminal-Bench 4.0. They are benchmark results, not evidence that Neon alone will be faster on a team's actual backlog.

That distinction is where the launch is useful. Model routing is not automatically a cost saving: a planner can add calls, handoffs and retries, and list price does not capture tool execution or the reviewer's time. Neon's place in the lineup is a proposed control for trading depth against speed; the value depends on the task, the route and the definition of “done.”

Test the patch, not the pitch

Run a small, fixed set of representative repository tasks through Neon alone, the routed path and your current baseline. Keep the checkout, permissions, tools and acceptance tests constant. Count changes that pass tests and survive human review; also log wall-clock time, retries, interventions and total model spend. Record cost per verified change, not just tokens per attempt.

Start with tasks that have observable outcomes—dependency updates, a reproducible bug, a contained refactor—and keep a reviewer accountable for the merge. If routing saves quota but creates more repairs, it may be a bad trade. If it repeatedly shortens time to an accepted patch without raising correction load, expand the trial. The launch makes orchestration a core product promise; the engineering team still needs evidence from its own toolchain.

Sources and editorial method

This AI-assisted report synthesizes the sources below. Automated source checking is not independent human verification; follow the original reporting for context and updates.

Common questions

How should a team evaluate model routing?

Compare the routed workflow with a single-model baseline on fixed repository tasks, measuring verified changes, review effort, time, interventions and total spend.