A coding model can finish a pull request in two minutes. If an engineer spends twenty minutes finding the missing test and undoing an unnecessary refactor, how fast was the model?
That’s a useful question to ask about Gemini 4 Neon, which Google introduced on November 18. Google positions it between Hydrogen for everyday tasks and Carbon for deeper, longer-running work. Neon has a one-million-token context window, four thinking levels, and is rolling out through the paid Gemini API and Google Antigravity on eligible plans. Google’s launch announcement · Antigravity release details
Start with a pull request your team already fixed
Pick a recently resolved bug that required changes across a few files. Give Neon and your current coding model the same issue description, repository snapshot from before the fix, and acceptance criteria. Keep the merged patch out of the materials they can see. Use it afterward as a reference, while allowing for other valid solutions.
A good test case might be a settings change where existing users must keep their preferences and new users should receive a different default. The task is easy to describe, but a patch can look correct while quietly changing behavior for existing users.
Give both outputs the same review you’d give a teammate’s pull request.
| Check | What to record |
|---|---|
| Correctness | Did the patch fix the issue and preserve existing behavior? |
| Scope | Did it limit changes to the files and behavior the task required? |
| Review effort | How many minutes did a developer spend checking and correcting it? |
| Cost | What did the full attempt cost, including retries? |
Try a change of direction midway
Real requirements change. Once each model proposes a fix, add a constraint: the old configuration format must remain readable for one more release.
See whether the model updates its plan, revises the relevant tests, and preserves the work that still applies. This tells you something a clean first attempt can’t: how much rework the model creates when you steer it halfway through a task.
You can also run the same issue at two thinking levels. A deeper setting earns its place if it catches a missed compatibility case or cuts review time enough to justify the extra latency. For simpler changes, the faster setting may already meet your acceptance criteria.
Decide where Neon fits
Try this with a small mix of tasks: one bug fix, one refactor, and one change that crosses module boundaries. Compare the total time from issue to review-ready patch. If Neon helps most on cross-module changes, that’s a good place to start using it. You can keep measuring as new tasks come in.
What would make you switch coding models first: less review time, fewer regressions, or lower cost per completed task?
