Gemini 4 Neon in Copilot: What Its Benchmarks Don't Say
Gemini 4 Neon is GA in GitHub Copilot. Its 61 index score is at xhigh, FrontierCode 46.2% at high, and SWE-sweep swings 12.8% to 39% on harness.

Gemini 4 Neon is now generally available in GitHub Copilot, and its headline benchmark scores were each produced at a different setting. Before you switch your agent over, it is worth knowing which number came from where.
Where Neon sits
Neon is Google's middle model: it sits between Carbon, the flagship aimed at the hardest long-horizon work, and Hydrogen, the fast everyday option. The pitch is intelligence approaching Carbon at a fraction of the time and token cost, built specifically for long-running agents and multistep tool use.
In Copilot it is rolling out gradually to Pro+, Max, Business, and Enterprise. One deployment detail matters for admins: under default model enablement, anyone who has not disabled it will see Neon appear automatically.
The scores come from different settings
Neon scores 61 on the Artificial Analysis Intelligence Index, third overall — but at the xhigh thinking level. Its FrontierCode 2.0 result of 46.2% is at the default high setting. These are two configurations, not one, so a single quoted Neon score without its thinking level does not describe a model you can order.
The widest spread is on SWE-sweep: 12.8% with the mini-SWE harness versus 39.0% inside Google Antigravity. Same model, nearly three times the result, because the agent system around it carries much of the load. The 39% belongs to Antigravity's harness — not to Neon in a Copilot session.
On steerability, the capability that decides whether an agent follows your mid-task corrections, Neon ranks second on the new Agent Arena Index across more than 4,200 real-world agent sessions. The caveat attached to that ranking in the analysis is that Google has not published the detailed methodology, so it reads as promising rather than proven.
Pricing, and what the percentages leave out
Token prices are $3 per million input, $0.15 per million cached input, and $15 per million output. Google claims 32% lower cost per task than Carbon at the high setting, and 50-80% lower than top competing frontier models on some indexes.
Those percentages are vendor math. Cost per task depends on how much work a task consumes and whether the output is usable on the first pass, and the announcement does not provide enough underlying detail to convert a percentage into a forecast for a specific team's bill. The cached-input price is the figure most likely to move a real invoice for agent workloads, since long-running agents re-send context constantly.
What else ships with it
Neon carries a 1M-token context window and handles spreadsheets, PDFs, charts, audio, and video into structured outputs, with a reported 73.5% on Chartography for chart reading — an evaluation score, not a guarantee across arbitrary documents. Inside Antigravity it can orchestrate parallel Hydrogen subagents, a documented workflow in that product rather than a capability every surface exposing Neon will offer.
Is it worth switching to?
If your agents run inside Copilot, the useful comparison is not Neon versus Carbon on an index — it is Neon versus whatever you run today, measured on your own task set, in your own harness. The SWE-sweep gap is the strongest argument in the launch data for doing that: a 3x swing from harness alone will dwarf most model-to-model differences you are trying to detect.
The practical move is to keep a fixed set of tasks you have already run, replay them on Neon at a thinking level you can afford, and compare both output quality and actual spend. Benchmarks at xhigh will not tell you what high costs you in production.
More from DangMua