A new model in the picker is tempting to try on the biggest problem in the repo. That is exactly the wrong first job for Gemini 4 Neon. GitHub made Neon generally available in Copilot on November 18, 2026, and Google built it for agentic coding and work that runs across many steps and many tool calls. A model that can keep going needs a smaller fence for its first run, not a bigger one.
This tutorial sets that fence up in order. You will check that your plan can see Neon at all, confirm the model policy on a Business or Enterprise account, select Neon in Visual Studio Code, and run it first in Ask mode, where it answers and edits nothing. Then you will move to one small agent task on a branch you can throw away, with approvals left on and a diff you read line by line before you keep anything.
What You Will Build or Learn
You will end with Neon selected in Copilot Chat in Visual Studio Code, a fresh branch ready for a first test, and a simple routine you can repeat every time a powerful new model lands in your picker. The routine is the point. Models will keep arriving in Copilot, sometimes several in one week. A safe start routine means each one earns trust on work you can check, instead of inheriting trust from the last one.
You will also learn where the traps sit. Changing the chat model does not change the model behind inline suggestions, GitHub says in its docs. On a work account, an administrator can leave Neon off even when your seat is paid. And under default model enablement, GitHub says new models are enabled on their own unless an administrator turned off the global default or disabled the model itself. Both directions need a check before you blame your editor.
One branch for the first run. If the work goes wrong, the branch goes away and main never felt it. Illustration: Dirty Static
Before You Begin
Know what Neon is before you hand it work. Google says Neon is adapted from Gemini 4 Carbon and sits between Carbon and the fast Hydrogen model. Its model card lists text, image, audio, video and PDF input with text output, a context window of about one million tokens, and a knowledge cutoff of March 2026. Google reports quality close to Carbon at a lower token cost per task, faster first responses, and fewer unrequested changes on coding tasks. Those are Google claims, relayed by GitHub. No independent result in the sources for this tutorial proves them on your repo.
Price needs the same plain look. GitHub bills Neon at provider list pricing under usage based billing. Google lists the Neon API rate at $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens. A long agent session resends a lot of context across many turns, so the bill for a sprawling first task can surprise you even when the token rate looks modest. Small first, then scale.
Step 1: Check Your Plan and the Model Policy
Action: confirm your Copilot plan first. Neon is available to Copilot Pro+, Max, Business and Enterprise users, per the GitHub changelog. If you are on Free or Pro, stop here. The picker will not show Neon, and no editor setting fixes a plan limit.
On a Business or Enterprise account, ask an administrator to check the model policy in Copilot settings before you troubleshoot anything local. GitHub says administrators manage access to Neon through that policy. Under default model enablement, new models are automatically enabled unless an administrator turned off the global default or explicitly disabled this model. So on most work accounts Neon should appear on its own, and when it does not, policy is the first suspect after rollout timing.
Expected result: you know your plan covers Neon, and on a work account you know the policy allows it. Write both facts down for your team. The next model launch will raise the same two questions.
Step 2: Update Your Client and Pick Neon in VS Code
Action: update Visual Studio Code and the GitHub Copilot extension, then open Copilot Chat and select the model, following the steps in GitHub documentation for changing the chat model.
- Click the chat icon in the title bar to open Copilot Chat.
- At the bottom of the chat view, open the CURRENT-MODEL dropdown.
- Select Gemini 4 Neon from the list.
Expected result: the dropdown at the bottom of the chat view now names Gemini 4 Neon. If you were using auto model selection, GitHub docs say to start a new chat session to switch models, so open a fresh chat after the pick. If Neon is missing from the list, do not reinstall at random. Rollout is gradual, GitHub says, so an eligible account may simply not have it yet. Check back later, and check the policy from Step 1 again on a work account.
One limit matters here. This pick changes the chat model only. GitHub docs state that changing the model used by Copilot Chat does not affect the model used for inline suggestions. Your tab completions keep running on their own model and their own setting, whatever you pick for chat.
Step 3: Start in Ask Mode in a Fresh Chat
Action: in the new chat, leave the mode on Ask for your first exchange. Ask mode answers questions and does not edit files or run commands as an agent task does. Use it to make Neon prove it understands the repo before it touches the repo.
A good first prompt names one file, one behaviour and one limit, in plain words. For example, ask Neon to explain how a named module handles errors, and to name the files it read to reach that answer. Read the answer against the code yourself. A wrong explanation costs you a minute. A wrong agent edit on the same misunderstanding can cost you an afternoon.
Expected result: Neon gives an account of the code that matches what you see in the files it names. If the answer invents a function, a file or a test result, stop. GitHub relays a Google claim that Neon is less likely to make unsupported claims about code, tool or test results. Less likely is not never, and your Ask round is where you find out how it behaves on your code, at no risk to the code.
Step 4: Make a Clean Branch for the First Agent Task
Action: in the terminal in Visual Studio Code, check that your tree is clean and create a branch for the test. Commit or stash any work of your own first if the status check shows changes.
git status
git switch -c neon-first-test
Expected result: the status check shows no changes you care about, and the new branch is checked out. Every edit Neon makes now lands on neon-first-test. You can read the full change with git diff, and you can drop the branch whole if the work is poor. Main stays untouched either way.
Step 5: Give Agent Mode One Small Bounded Task
Action: switch the chat to Agent mode and give Neon one task with a hard edge. Pick work that is small, covered by tests, and easy to judge: a named bug with a known failing test, a rename in one module, or a missing validation check in one function. State the files it may touch, the test it must pass, and the things it must not change, in the prompt itself.
Keep approvals on for this run. When Copilot asks to run a terminal command or use a tool, read the request before you approve it. A command that installs a package you never named, reaches for the network for no clear reason, or deletes files is a stop sign, not a detail. Decline it and narrow the prompt.
Expected result: the task finishes inside the files you named, or it stops and explains why it cannot. Both outcomes are useful data. A clean stop on a hard edge tells you more about working with this model than a sprawling success you cannot verify.
Our review of GitHub Copilot Browser Use covers the same approval habit in a browser setting, where an agent acting inside tabs needs the same kind of fence.
Step 6: Read the Diff, Run the Tests, Check the Bill
Action: before you keep any change, read the full diff yourself. Then run the tests for the area the task touched, yourself, in your own terminal. Do not take the chat summary as proof that a test passed. The summary is a claim. The test run is evidence.
git diff
git diff --stat
Expected result: the diff shows only the change you asked for, in the files you named, with no extra edits riding along. The --stat view gives you the shape of the change at a glance, which makes an unexpected file easy to spot. Your own test run passes, or you send the failure back with the exact output and let Neon try once more on the branch.
Finish with the money check. Open your Copilot usage view and find what the session cost under usage based billing. Compare it with the size of the task. That ratio, cost against a small known job, is the number that tells you whether Neon at $3 in and $15 out per million tokens earns a place on bigger work in your repo. The cost per task claim Google makes for Neon against Carbon means little until your own tasks put a figure beside it.
The diff is the review. The chat summary is only the claim about the review. Illustration: Dirty Static
Verification
Run these checks and record the results. The chat dropdown names Gemini 4 Neon in a fresh session. Ask mode gave an account of a named module that matched the files it cited. The first agent task stayed on the neon-first-test branch and inside the files the prompt named. Your own test run, not the chat summary, shows the result. The usage view shows what the session cost, and you can state that cost beside the task it paid for.
If you can tick each line, Neon is set up in Copilot in a way you can defend to a teammate or a manager. If any line fails, the Troubleshooting table below names the likely cause.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| Neon is missing from the model picker | Rollout is gradual, or the plan does not cover Neon | Confirm a Pro+, Max, Business or Enterprise plan, update the client, and check back later, as GitHub advises |
| Neon is missing on a work account only | The model policy in Copilot settings leaves it off, or the global default is off | Ask an administrator to check the model policy for Neon, per Step 1 |
| The chat still answers with the old model | Auto model selection was on when you switched | Start a new chat session after picking Neon, per GitHub docs |
| Inline suggestions did not change | Chat and inline suggestions use separate model settings | Change the completions model in its own setting if you need it changed at all |
| The agent touched files outside the task | The prompt named no file limit, or approvals were waved through | Restore the branch state, narrow the prompt to named files, and keep approvals on |
| The chat says tests pass but your run fails | The summary reported a claim, not your evidence | Trust your own run, paste the exact failure output back, and let the branch absorb the retry |
| The usage bill is higher than expected | A long agent session resends context across many turns at provider list pricing | Start future work in smaller tasks, and compare cost per task on known jobs before scaling up |
Next Steps
Keep Neon on bounded work for a week of real tasks, and keep notes on two things: how often the diff needed a correction, and what each finished task cost. Those notes beat any launch chart when you decide what Neon should own in your repo. When you are ready to tune how hard it thinks on each job, our guide to thinking levels in Gemini 4 Hydrogen covers the same four level idea in the sibling model, and the Neon round of that guide follows in the next Tutorials slot. For a contrast in price and pitch, the MAI Code 2 review shows what a model built to win on cost per task looks like inside the same Copilot picker.
The habit worth keeping is the order, not the model name. Plan, policy, picker, Ask, branch, small task, diff, tests, bill. Run that order for every new model and the picker stops being a shelf of temptations. It becomes a set of tools, each with a first job it actually passed.
Facts checked against the GitHub changelog for Gemini 4 Neon in Copilot and GitHub documentation on changing the Copilot Chat model on November 19, 2026. Model facts and prices are Google published values and can change. This procedure was verified against those documents and is not presented as a personal test run.
Sources
- GitHub, Gemini 4 Neon in GitHub Copilot, changelog, November 18, 2026: https://github.blog/changelog/2026-11-18-gemini-4-neon-in-github-copilot/
- GitHub Docs, Changing the AI model for GitHub Copilot Chat, checked November 19, 2026: https://docs.github.com/en/copilot/using-github-copilot/ai-models/changing-the-ai-model-for-copilot-chat?tool=vscode
- Google DeepMind, Gemini 4 Neon model card, published November 18, 2026, via launch coverage checked November 19, 2026
- Google, Introducing Gemini 4 Neon, November 18, 2026, pricing and thinking level details via launch coverage checked November 19, 2026
- Hero, branch and review images: Dirty Static illustrations
END OF TRANSMISSION
Stay curious. Question everything.


