Gemini 4 Neon: The Middle Child of the Gemini 4 Family
Wednesday morning began with a leak. A price list inside Google Cloud’s billing catalogue quietly named a model nobody had announced — Gemini 4 Neon, at $3 per million input tokens and $15 per million output tokens. By the evening, the rumour had become a launch, and Google’s AI family had a fourth member in four weeks.
Koray Kavukcuoglu, senior vice president at Google DeepMind and Google’s chief AI architect, introduced the model on November 18, calling it “frontier-level intelligence with fast end-to-end task completion.” The pitch is aimed squarely at people who do complex work every day — software engineers, analysts and researchers — and who are tired of choosing between speed and power.
To understand Neon, you need the family photo. Argon arrived in late September as the flagship. Gemini 4 Carbon followed in late October as the model for the hardest problems. Hydrogen landed on November 12 as the fast, cheap workhorse for everyday tasks. Neon, announced on November 18, sits deliberately in the middle.
Google describes Neon as a fast reasoning model. It is built for complex software engineering, knowledge work and scientific applications — the kind of jobs that are too demanding for a lightweight model but do not always need the heaviest gun in the arsenal.
The company says Neon’s quality comes close to Carbon on coding, reasoning and agentic tasks, but at faster speeds and lower cost. It builds on Hydrogen, Google says, with deeper reasoning added on top. In the Gemini app’s model picker, the choice is now plain: Everyday help runs on Hydrogen, Deeper analysis runs on Neon, and Code and complex work runs on Carbon.
Gemini 4 Neon Pricing: Cheaper Per Task, Not Per Token
Neon’s price list is where the story gets interesting, because it looks contradictory at first glance. Developers pay $3 per million input tokens and $15 per million output tokens. Cached input costs $0.15 per million tokens — a 95 percent discount that rewards long, repeated conversations.
Those numbers are exactly double Hydrogen’s $1.50 and $7.50. And through the end of November, Carbon and Argon are actually cheaper per token, at an introductory $2 and $10. On a price list alone, Neon looks like the expensive middle child.
Google’s answer is that a price list is not a bill. At its default “high” thinking level, the company says Neon costs 32 percent less per task than Carbon — and roughly 50 to 80 percent less per task than top competing frontier models at their highest settings. The savings, Google argues, come from finishing jobs in fewer steps and with less wasted effort.
The calendar matters too. Carbon’s introductory pricing ends on November 30. From December 1, Carbon and Argon rise to $4 per million input tokens and $20 per million output tokens. From that day on, Neon undercuts them by 25 percent per token as well as per task. Anyone budgeting for December should compare the post-November prices, not today’s.
There is one honest caveat. Google has published no benchmark scores of its own for Neon yet. The model card and documentation describe what the model can do, not how it scored. The “close to Carbon” claim, for now, rests on Google’s own framing rather than independent numbers — though one independent index has already weighed in.
How Fast Is Gemini 4 Neon? Benchmarks and Scores
The Artificial Analysis Intelligence Index gives Neon a score of 61 at its highest thinking level, placing it third overall — just behind larger, slower models. It also takes the number two spot on the Vals Index, with particular strength in legal research, finance, accounting and other analytical work.
Speed is the other half of the claim. Google pitches Neon at Flash speeds — the quick, responsive feel of its lightweight models — while keeping performance near Carbon’s level, according to 9to5Google. It has the same 1-million-token context window as its siblings, with four thinking levels to choose from: Low, Medium, High and Extra High.
The thinking levels are the dial that makes the pricing work. A quick question can run at Low and cost almost nothing. A tangled coding problem can run at Extra High and still, Google says, come in cheaper per completed job than the flagship would. There is also a cyber-tuned variant of Neon, built for security work, which will reach eligible members of Google’s Fairwind Program — the initiative that gives cyber defenders early access to frontier models — over the coming days.
Who Gets Gemini 4 Neon First
Neon is rolling out now to Google AI Pro and AI Ultra subscribers in the Gemini app. That detail matters more than it sounds. Until now, the strongest Gemini models were reserved for Ultra, the most expensive tier. Pro subscribers — who pay considerably less — now get a frontier-class model in their picker.
With the launch, Neon replaces the older 3.1 Pro model in the app, and it also takes Argon’s place in the picker for most users. The message is clear: Google wants its mid-tier subscribers using a model that can handle serious work, not just quick questions.
Developers can reach Neon through the Gemini API, in Google AI Studio and in the Vertex AI Model Garden. Paid subscribers can also use it inside Google Antigravity, Google’s agentic development environment, where it takes on a special role.
Gemini Tasks — the feature that turns the assistant into something that works through multi-step jobs on your behalf — can now run on Neon as well, working through complex tasks more quickly and efficiently than the larger models, according to Google. If you use the Gemini app, it is also worth knowing what Google does with your chats — the company recently split its privacy controls so you can keep history while opting out of AI training.
Gemini 4 Neon in Antigravity: The Orchestrator Role
Inside Antigravity, Neon is not just another model on a menu. Google describes it as a fast, high-judgment Orchestrator — a model that plans a job, breaks it into pieces and directs parallel Hydrogen subagents to carry them out.
That division of labour is the point. A single powerful model doing everything is slow and expensive. A smart coordinator directing a fleet of cheap, fast workers can be both quicker and cheaper — if the coordinator is good enough. Neon is Google’s bet that it is, and if it works, the cost of building software with AI agents falls sharply.
What Gemini 4 Neon Means for Google’s AI Strategy
Four models in four weeks is a pace, not an accident. Google is slicing its lineup into ever finer tiers — Argon and Carbon at the top, Neon in the middle, Hydrogen at the bottom — so that every kind of work has a model priced for it, and every kind of customer has a reason to stay inside Google’s ecosystem.
The strategy also answers a question that has followed Google all year: how do you sell frontier intelligence without giving it away? Neon’s answer is to sell it by the task, not by the token — and to let the cheaper Pro plan carry a model that, until this week, would have been an Ultra exclusive. For users, the practical effect is simple: the model picker now reads like a menu with honest labels, and most people will spend most of their time in the middle. That is exactly where Google wants them.
What Happens Next
The first test will be the independent benchmarks. Google’s own numbers — 32 percent cheaper per task than Carbon, third on the Artificial Analysis index — are promising, but developers will want to see Neon put through real coding and reasoning tests by third parties before they rebuild their workflows around it.
The second test is December 1. When Carbon’s price rises to $4 and $20, Neon’s position as the sensible middle tier becomes much clearer — and Google will find out whether developers choose it for the price, the speed, or both. And the third test is the one Google cannot control: whether rival labs answer with their own mid-tier models. OpenAI, Anthropic and the open-weight labs have all been racing in the same direction, so Neon is unlikely to be the last mid-tier model of the year.
For now, though, the leak that started the day has become the launch that ends it — and AI Pro subscribers have a new model in their pocket that was, until Wednesday morning, not supposed to exist yet.



