Google Gemini 4 Neon arrived on 18 November 2026, and it arrived fast in more than one sense. It is the fourth model in the Gemini 4 family to launch in four weeks, following Carbon, Hydrogen and Argon, and Google describes it as offering "frontier-level intelligence with fast end-to-end task completion" for software engineering, knowledge work and science. That description is doing two jobs at once: it places Neon at the top end of capability, and it says the point of the model is speed through a whole task, not simply a good answer to a single question. For anyone trying to follow a launch cycle that has produced a new model almost every week, Neon is the one Google is pitching as the fast frontier option.

The headline numbers give the claim some shape. Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, which puts it third overall. It ranks first on the Coding Agent Index with a score of 58.6. It carries a 1 million-token context window, which is the amount of text, code or data the model can hold in view at once. It offers four thinking levels — Low, Medium, High and Extra High — so a user can trade speed against depth depending on the task. And its pricing is set at 3 dollars per million input tokens and 15 dollars per million output tokens, with cached input at 15 cents per million, a 95 per cent discount on the input rate. Those figures are the raw material for judging the model; the rest of this article explains what they mean in practice.

Google Gemini 4 Neon: what the benchmarks actually say

Benchmarks are a useful starting point if they are read with care, and Neon's two headline scores reward that care. A score of 61 on the Artificial Analysis Intelligence Index at xhigh thinking places Neon third overall. That is a strong general position, and it says the model competes at the top of the field rather than in a niche. The more specific score is the Coding Agent Index, where Neon ranks first with 58.6. A first-place ranking on a coding measure matters because coding is one of the main uses Google names for the model, alongside knowledge work and science. A model that leads on coding and sits third overall is a model whose strength is weighted toward building software.

The thinking levels are part of how those scores are reached. Neon offers four levels — Low, Medium, High and Extra High — and the 61 on the Intelligence Index is recorded at xhigh thinking, the deepest setting. That detail matters because a score at the deepest setting is not the same as a score at the default setting. A user who runs Neon at Low thinking for a quick task will get a faster answer with less depth, and the benchmark figure does not describe that run. The four levels are a control the user can set, not a single fixed behaviour, and the right way to read the benchmark is as a statement about what the model can do when it is allowed to think hardest.

The 1 million-token context window is the other number that shapes what Neon can do. A token is a small unit of text, and a million of them is a large working space: enough to hold a substantial codebase, a long document set, or a research file alongside the question being asked. For software engineering, that means the model can see more of a project at once. For knowledge work, it means more source material can sit in the same run. The window does not by itself make the model smarter, but it removes one of the limits that forces users to split large tasks into smaller pieces.

Google Gemini 4 Neon pricing: what it costs to run

Neon's pricing is set at 3 dollars per million input tokens and 15 dollars per million output tokens, with cached input at 15 cents per million tokens. Cached input is text the model has already seen and stored, so a repeated prompt can reuse it at a much lower rate — the 95 per cent discount on the input price. The gap between input and output pricing is normal for this kind of model: output tokens, the text the model generates, cost more than input tokens, the text it reads. A user running a coding task will pay more for the code the model writes than for the code it reads, and the cached-input rate rewards workflows that reuse the same context across runs.

Google's own comparison with Carbon, the model Neon follows, puts the cost case in plain terms. Google says Neon matched or exceeded Carbon on quality at three times the speed and 60 per cent of the cost, a claim attributed in Google's post to Thai Tran, the AI product lead at Glean. At High thinking, Neon costs 32 per cent less per task than Carbon. Those are Google's figures, and they are framed as a comparison with the model Neon sits alongside rather than with the whole market. The point of the comparison is that Neon is being sold as a cheaper and faster way to get Carbon-level quality, not simply as a more capable model at any price.

The market comparison comes from VentureBeat, which notes that Neon is 25 per cent below the standard rates for Anthropic's Opus 5.5, priced at 4 dollars per million input tokens and 20 dollars per million output tokens. That puts Neon's 3 and 15 dollar rates in context: below one of its main rivals on list price, though list price is only one part of what a task costs in practice. The thinking level chosen, the amount of cached input reused, and the length of the output all change the final bill. A user comparing models on price alone should run the same task through each and compare the real cost, not just the headline rates.

Who gets Google Gemini 4 Neon and what changes for existing users

Neon is rolling out in the Gemini API, through Google AI Studio and Android Studio, in Gemini for business, in Google Antigravity for paid Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. That list covers developers, business users and paying consumers, but it does not cover everyone: no free-tier access has been announced. A free user who wants to try Neon will need to wait for a free tier to be offered or move to a paid plan, and Google has not said when or whether that will happen.

For existing users, the launch carries two changes. Neon replaces Argon, the model that launched just before it in the same four-week run, which means Argon steps aside as the current option. And 3.1 Pro is retired for Pro and Ultra subscribers, which means subscribers who were using 3.1 Pro will need to move to a Gemini 4 model. The pace of that change is part of the story: four models in four weeks, a replacement within the same month, and a retirement for paying subscribers. Google is moving its model line quickly, and users who build workflows around a specific model will need to track which one is current.

There is one more rollout worth noting. A cyber-defence version of Neon is rolling out to trusted defenders under the Fairwind Program. That is a separate track from the general release, aimed at security teams rather than general users, and it shows Google positioning the model for defensive security work as well as for coding and knowledge tasks. The Fairwind Program is the channel for that version, and access is limited to the trusted defenders in the program rather than open to all.

Why Google Gemini 4 Neon matters and what happens next

Neon matters because of where it sits in the Gemini 4 family. Carbon, Hydrogen and Argon arrived first, each in its own week, and Neon arrives as the fourth in four weeks with a claim to match Carbon's quality at three times the speed and 60 per cent of the cost. If that claim holds in real use, Neon becomes the practical choice for tasks where speed and cost matter as much as raw capability: coding agents that run many steps, knowledge work that processes large documents, and science work that needs long context. The first-place ranking on the Coding Agent Index and the third-place ranking overall support that positioning, but the test that counts will be the work users run through it in the weeks after launch.

What happens next is a matter of watching three things. First, whether the speed and cost claims hold outside Google's own comparisons, once developers run Neon on their own codebases and compare the bills. Second, whether the retirement of 3.1 Pro and the replacement of Argon cause any friction for subscribers who had settled on those models. Third, whether a free tier appears, and how wide the rollout becomes beyond the API, AI Studio, Antigravity and the Gemini app for paying subscribers. The launch is one step in a fast-moving sequence, and the next step will be the evidence from real use. For readers tracking the wider AI market, the Nvidia third-quarter earnings report, the coverage of Jensen Huang's comeback talks, and the Anthropic IPO story give the market context around this release.

Sources: Google blog, 9to5Google, VentureBeat, RRTechPulse — November 2026.