Google has a new middle child, and it might be the most useful model in the Gemini 4 family. Launched on November 18, 2026, Gemini 4 Neon sits between the fast, cheap Hydrogen and the slow, deep Carbon — and according to Google, it delivers Carbon-class intelligence at roughly half the cost per task. For developers, knowledge workers and anyone building AI agents, that trade-off is the whole story.
Neon is not a minor refresh. It is the fourth Gemini 4 model Google has shipped in about seven weeks, and it replaces Gemini 3.1 Pro in the Gemini app while also taking over from the Argon model. If you pay for Google AI Pro or Ultra, Neon is already in your model picker. Here is what it actually does, what it costs, and whether it is worth switching to.
What Is Gemini 4 Neon?
Gemini 4 Neon is Google’s mid-tier frontier model, built for complex software engineering, knowledge work and science workflows that need to finish fast. Google describes it as adapted from Carbon, its most powerful model, but tuned for speed and efficiency rather than maximum depth.
The headline specs are straightforward: a 1 million-token context window, four thinking levels — low, medium, high and xhigh, with high as the API default — and multimodal input across text, images, audio and video. Neon can pull all of those inputs together in a single context window and turn them into structured deliverables, from code to spreadsheets to research summaries.
What makes Neon different from a typical mid-tier model is the emphasis on agentic work. Google highlights Neon’s steerability — how cleanly it adapts when you correct it mid-task — where it ranks second on Arena’s new Agent Arena Index with a score of 1336 across more than 4,200 real-world agent sessions. In plain terms, Neon is built for the back-and-forth of real agent workflows, not just single-shot answers.
Gemini 4 Neon Benchmarks: Where It Actually Wins

Vendor benchmarks always deserve a sceptical eye, but Neon’s numbers are unusually strong for a mid-tier model — and several come from independent evaluators. Google reports that Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, ranking third overall and just behind larger, slower models. On the Vals Index it ranks second with 70.9%.
The coding results are the most eye-catching. Neon leads the Artificial Analysis Coding Agent Index at 58.6, scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting. For science and maths, Google says Neon is best in class on Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%) — the same kind of hard reasoning that made headlines this week with Claude’s claimed proof of the Riemann Hypothesis.
Neon also takes first place on both the Legal Research Bench and the Tax Agent Bench, leads most Surge AI categories across finance, accounting and spreadsheet modelling, scores 73.5% on Chartography for reading dense professional charts, and wins Zapier’s AutomationBench (56.3%) and Andon Labs’ Vending-Bench 2 — a long-horizon simulation of running a business — where it earned $16,843.
One caveat is worth stating plainly: the headline 61 on the Artificial Analysis index was measured at xhigh, a thinking level app users cannot select. At the default high level, independent reviewers put Neon at 59.6 — still level with Claude Opus 5.5 at its maximum effort. Treat Google’s rankings as vendor-reported until you see them confirmed on the third-party leaderboards, but the independent coding tests so far back the story up.
Gemini 4 Neon Pricing: The Real Cost per Task

Neon is priced at $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million — a 95% discount that matters a lot for agents that resend the same large context on every turn. That is twice Hydrogen’s price, but Google says Neon costs 32% less per task than Carbon at the default high level, and 50% to 80% less per task than top competing frontier models on the Vals and Agent Arena charts.
There is a wrinkle in the headline price. Neon writes a lot of tokens, so the per-token price is only half the story. Independent testing found Neon costs about $3.74 per task at the high setting, against $7.76 for Carbon — roughly half — while the xhigh setting adds about $2.70 per task for 1.8 extra index points. The practical advice from reviewers is to run Neon on high, not xhigh, unless you need the last drop of capability.
Context also matters: Carbon’s list price is $4 input and $20 output, but it is on promotional pricing of $2 and $10 until November 30. After that, Neon ends up 25% cheaper than Carbon on list price. Hydrogen, at $1.50 and $7.50, remains the budget option for everyday tasks.
Where You Can Use Gemini 4 Neon Today
Neon rolled out on launch day in the Gemini API in Google AI Studio, in Gemini for business, in Google Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. Free users do not get it yet — it is a paid-tier model for now.
That matters because Neon is arriving just as AI assistants are spreading across every surface. Signal’s new screenshot blocking shows how seriously the industry is taking the privacy side of OS-level AI, while Apple’s AI agent that runs your Mac and Microsoft’s always-on Copilot Autopilot show agents moving from chat windows into the operating system itself. Neon is Google’s answer for the model that powers that kind of always-on, multi-step work — fast enough to feel instant, smart enough to be trusted with the job.
For developers, the most immediate use is coding. Neon is the upgrade target Google points Gemini 3.1 Pro users toward, and for complex coding work moving off Argon or Gemini 3.8 Flash. If you run agents that write, test and fix code across a large codebase, the 1 million-token context window is the feature that changes the workflow: the whole repo can stay in context while the agent works.
9to5Google’s coverage notes Neon is available in the Gemini app for AI Pro and Ultra subscribers.
Gemini 4 Neon vs Carbon vs Hydrogen: Which One Should You Pick?
The Gemini 4 family now has a clear ladder, and the choice is simpler than the naming suggests.
Hydrogen is the fast, cheap one — $1.50 in and $7.50 out per million tokens — for everyday execution and quick answers. Carbon is the slow, deep one for your hardest, long-horizon projects, at $4 and $20 on list price. Neon sits in the middle: Carbon-adjacent intelligence at Flash-like speed, for complex coding, analytical reasoning and production-grade enterprise workflows.
The numbers back the positioning. Against Carbon, Neon matches the Artificial Analysis index score at high effort for roughly half the cost per task and about 60% less time. Against Claude Opus 5.5, Neon at high is level on the index at about a third of the per-task cost. Against GPT-6.2 Sol and Astra, Neon wins on long context and the Vals Index but loses on cost per task. And if you want to run frontier AI without the cloud at all, Nvidia’s RTX Spark laptops that run frontier AI locally are the hardware side of the same trend.
One more consideration: Google says Neon records its lowest fabrication rate yet in internal Antigravity testing, continuing work to curb hallucinations. The full safety report is still on the way, so if you work in a regulated space, wait for that before putting Neon into production.
Who Should Switch — and Who Should Wait
Switch to Neon now if you are paying for Gemini 3.1 Pro, running coding agents on Argon or Gemini 3.8 Flash, or building agents that juggle long contexts and mid-task corrections. It is the default sensible choice in the Gemini 4 line-up, and the price makes it easy to justify.
Wait if you only need quick everyday answers — Hydrogen is cheaper and fast enough — or if your work demands the absolute deepest reasoning, where Carbon still leads. And as with any model that has been out for three days, test it on your own tasks before committing: it is the fourth Gemini 4 model in seven weeks, and the benchmarks, however strong, are not your workload.
The Bottom Line
Gemini 4 Neon is Google’s bet that the most valuable AI model is not the smartest or the cheapest, but the one that finishes hard jobs quickly at a fair price. Carbon-level coding and reasoning, a million-token context window, best-in-class steerability for agents, and half the cost per task of its bigger sibling — that is a combination worth taking seriously. The mid-tier just became the sweet spot.

