Models

Gemini 4 Neon Lands in the App, and the Model Picker Loses Two Names

Google's fourth Gemini model in a month slots into the middle of the lineup, priced to be used every day. The app now asks subscribers to choose by effort, and two familiar names are gone.

Daniel HarrisDaniel Harris
Heat: 1,500
Three neon rings orbiting a small glowing core on a deep blue field

Dek: Google's fourth Gemini model in a month slots into the middle of the lineup, priced to be used every day. The app now asks subscribers to choose by effort, and two familiar names are gone.

What Gemini 4 Neon changes for anyone paying for the app

Google DeepMind published "Introducing Gemini 4 Neon" on 18 November, and the post reads less like a launch than a reshuffle. Neon doesn't extend the family at the top. It lands in the middle and pushes two older options out.

That's the part a subscriber feels. The model you learned to trust may not be on the list tomorrow morning.

Neon replaces Gemini 4 Argon in the app. According to 9to5Google, Gemini 3.1 Pro is retired at the same time, and 3.5 Flash-Lite becomes the "Auto" entry for paid users. Anyone who built a habit around either name has to move, and Google hasn't published a setting that keeps the old behaviour.

Retiring a model option doesn't delete your chats. Only the choice disappears.

The Gemini model picker now asks how hard the job is

The Gemini model picker used to work like a menu of capabilities, where the newest name was usually the right one. Google has turned it into a menu of effort instead, with four rungs and a price signal behind each.

Option

What Google says it's for

Auto

Everyday requests, routed by Google

Gemini 4 Hydrogen

Fast, everyday help

Gemini 4 Neon

Deeper analysis at Flash-like speed

Gemini 4 Carbon

Code and long-horizon work

Carbon stays at the top as the slow, expensive choice for the hardest jobs, the model Google measures everything else against. Hydrogen is the quick one you use without thinking. Neon is the middle rung, and Google's pitch is that it matches Carbon on most work while running at something close to Flash speed.

So which one do you pick for a 40-page contract on a Tuesday afternoon? Probably Neon, then Carbon if the answer looks thin.

What Gemini 4 Neon costs per million tokens

Pricing is the most concrete thing here, because a price is a number you can check against your own invoice.

Neon costs $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. Google calls the cached rate a 95 percent discount, and it's the one that matters if you paste the same long document into a chat twice a day. The context window is one million tokens.

Against the premium tier, that leaves Neon well under the top model's per-token cost, which is the point of a middle rung.

The benchmark numbers Google put in the post

Google's post is dense with scores. Four are worth keeping, if you read the attribution on each.

Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, third overall. The index bundles reasoning and knowledge tests into a single number, so third place means it's close to the leaders rather than ahead of them.

It ranks second on the Vals Index at 70.9 percent. That benchmark asks a model to finish multi-step work of the kind an analyst would hand off, and Google says Neon also sits on the index's Pareto frontier, the curve where no rival buys more accuracy for the money.

On the Artificial Analysis Coding Agent Index, which tests whether a model can finish a coding job handed to it by an agent, Neon ranks first at 58.6. Google reports 65.2 percent on Terminal-Bench 4.1 as well, a test drawn from real command-line tasks.

Neon also finished first on Andon Labs' Vending-Bench 2 with $16,843, a simulated vending business that runs for a long stretch and has to keep its books straight.

Vendor benchmarks, all of them. None of these has been rerun outside Google.

Gemini 4 Neon costs less per task than Gemini 4 Carbon

Scores win arguments on forums. Cost per task wins budgets, and this is the claim that will decide whether teams switch.

Google says Neon costs 32 percent less per task than Carbon at the default high thinking level. That's a different measure from price per token, because it includes the tokens each model burns to finish the same job. A cheap model that rambles can still lose to an expensive one that doesn't.

Cheaper per job.

The larger claim: Neon runs 50 to 80 percent cheaper per task than leading frontier competitors on the Vals and Agent Arena indexes, and at xhigh still costs at least 36 percent less than they do at their highest setting. Google attributes every figure to its own runs.

Neon also ships with four thinking levels, so a team can push accuracy up for a hard ticket and pull it back down for bulk work. Inside Google Antigravity, the paid coding environment, Google positions Neon as a fast orchestrator that can hand parallel jobs to Hydrogen subagents.

Where Gemini 4 Hydrogen fits, and why the middle rung matters

Three models doing three jobs is a cleaner story than four names on a dropdown, and it explains the retirements better than any benchmark.

Hydrogen handles volume. Neon handles work that needs a second look. Carbon handles the jobs where a wrong answer costs more than the compute.

The catch sits in the default. Google sets Neon to high thinking, not xhigh. A quiet admission that most requests don't need the top setting.

What Gemini AI Pro subscribers get on day one

Rollout is broad, but it isn't free. Neon arrives in the paid Gemini API tier, in Google AI Studio, in Gemini for business through the API and the model dropdown, and in Google Antigravity for paid Gemini AI Pro and Ultra subscribers.

In the Gemini app, it's Pro and Ultra subscribers only. The free tier is untouched, so nothing about this launch reaches people who never paid for a plan.

The Fairwind cyber model and the gaps Google left open

Google also built a cyber-tuned variant, and its numbers are the most striking in the release. It scores 76 percent on the hard tier of AutoBaxBench, 90.5 percent on SEC-bench Pro and 51.26 percent on WebExploitBench v1.2 at roughly half the cost of larger rivals.

You can't have it. The variant goes to vetted defenders through the Fairwind program, not to anyone with a credit card. That's a defensible call, and it also means the headline capability isn't testable today.

The rest of the gaps are ordinary. Google ran the benchmarks, Google ran the cost comparison, and The Decoder notes this is the company's fourth new Gemini model in a month, which makes lineup churn as much of a story as any single score.

If you're on a paid plan, the useful move is dull. Open the picker, run a task you know well at high and again at xhigh, then compare the week's bill. Cheaper, probably. Not weaker, as far as anyone outside Google can tell.

Share This Story

Sources

Related AI News

A terminal window with a rising reward curve and a 27B badge beside it
Models

RadixArk Fine-Tuned a 27B Model on 1,200 Tasks and It Beat the Original Everywhere

RadixArk took an existing open model, spent 18 days and about 1,200 hand-built tasks on it, and changed nothing about its shape. The gains are real, and the absolute scores are still low.

Heat: 1,330
A laptop outline with three stacked layer sheets passing through a loop arrow inside it
Models

Inkling 2 Mini Fits on a Laptop by Running the Same Layers Three Times

Thinking Machines Lab built a model that thinks longer by running the same block of layers three times. The trick buys reasoning without extra memory, and the bill comes due somewhere else.

Heat: 1,710
A long teal audio waveform with a single orange accent bar on a dark field
Models

A 0.5B Speech Model Just Matched a 30B One on Real Phone Calls

Reson8 shipped a half-billion-parameter transcription model that gets within a tenth of a point of a 30B rival on noisy phone calls. The interesting part is what that does to a per-minute bill.

Heat: 560
A large model core ringed with reasoning spokes beside a smaller fast core, with falling price bars on the right
Models

OpenAI's GPT-6.2 Astra Is 20% Cheaper and Still Not the Best at Everything

OpenAI cut the flagship's token prices and left the cheap model where it was. The leaderboards OpenAI doesn't run put the expensive one seventh.

Heat: 1,900