Gemini 4 Neon Explained: Google’s New $3 Coding Model That Tops the Agent Charts

Google has launched yet another Gemini 4 model, and this one arrived with a twist: its price leaked in Google Cloud’s own billing catalogue hours before the official announcement.

Gemini 4 Neon went live on 18 November 2026 as the fourth member of the Gemini 4 family, and it is aimed squarely at the work developers and knowledge workers actually do every day — complex coding, long documents, and multi-step agent tasks. Google pitches it as frontier-level intelligence with fast end-to-end task completion, and the early numbers back up at least part of that claim: Gemini 4 Neon tops the Artificial Analysis Coding Agent Index and costs a third less per task than the flagship Gemini 4 Carbon.

So what exactly is Gemini 4 Neon, how much does it really cost, and should you switch to it? This guide breaks down the benchmarks, the full pricing table, where you can use it today, and the honest caveats Google’s own model card quietly includes. Everything below is based on Google’s official pricing page, the DeepMind model card published on launch day, and independent leaderboard data — and where a number comes from Google itself rather than an independent test, we say so.

Developer coding at night with Gemini 4 Neon
A developer working late on code — exactly the kind of long, multi-file session Gemini 4 Neon is built for.

Table of Contents

What Is Gemini 4 Neon?

Gemini 4 Neon is Google DeepMind’s newest large language model, announced on 18 November 2026 by Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Chief AI Architect at Google.

According to the model card DeepMind published the same day, Gemini 4 Neon is adapted from Gemini 4 Carbon — the flagship of the family — but tuned for speed and cost rather than maximum capability. Google’s own description on its pricing page puts it plainly: Gemini 4 Neon “brings frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows, featuring a unified 1M-token context window and long-decode generation up to 1M tokens.”

That positioning matters because it tells you who the model is for. Gemini 4 Neon is not a chatbot model for casual questions, and it is not the cheapest model in the family either. It is the middle option: strong enough to handle repository-scale coding and long-horizon agent work, fast enough to feel responsive in an IDE, and priced so that running it all day does not sting. In Google’s words, it is built for “frequent, complex work” — the tasks you repeat dozens of times a week rather than the one-off moonshot you would hand to Carbon.

The launch itself was unusual. On the morning of 18 November, the model’s $3/$15 pricing surfaced in Google Cloud’s public billing catalogue before any announcement existed — thirty priced line items across the Gemini API and Vertex AI services, with no model card, no documentation page, and no endpoint to call. Hours later, Google made it official: the documentation page that had returned a 404 that morning now described gemini-4-neon in full, a model card was live on DeepMind’s site, and the model appeared in the Vertex AI Model Garden and Google AI Studio. The leaked price turned out to be exactly right.

Gemini 4 Neon accepts text, images, audio, video, and PDF files as input, and it replies in text. It does not generate images or audio, and it does not support the Live API, so this is a workhorse for agents and long tasks rather than a multimodal showpiece. Its knowledge cutoff is March 2026, which means events after that date — including its own launch — are outside what it was trained on, and Google warns that some domains may be less current than others.

Gemini 4 Neon at a Glance

Specification Gemini 4 Neon
Model ID gemini-4-neon
Launched 18 November 2026
Built from Adapted from Gemini 4 Carbon
Context window 1,048,576 tokens, shared across input, thinking, and output
Default output Up to 65,536 tokens per synchronous response; up to 1M tokens with long decode
Thinking levels Low, medium, high (default), xhigh
API price $3.00 per 1M input tokens; $15.00 per 1M output tokens
Cached input $0.15 per 1M tokens (a 95% discount)
Free tier None — paid tier only
Knowledge cutoff March 2026
Inputs Text, images, audio, video, PDF
Output Text only
Where to use it Gemini API, Google AI Studio, Android Studio, Gemini for business, Gemini app (Pro and Ultra), Antigravity, GitHub Copilot, Cursor, Vercel AI Gateway, OpenRouter

Two details in that table deserve a second look. First, the 1M-token context window is unified: input, thinking, and output all draw from the same 1,048,576-token budget. That is generous for whole-repository work, but it also means a long chain of thought eats into the room left for the answer. Second, there is no free tier at all. Every token Gemini 4 Neon processes is billed, including the thinking tokens it spends before it starts answering — a point Google states explicitly on its pricing page.

Where Gemini 4 Neon Fits in the Gemini 4 Family

Gemini 4 Neon is the fourth Gemini 4 model in about five weeks, and the family now has a clear shape. Each member has a distinct job, and understanding the lineup is the fastest way to decide whether Neon is the right pick for you.

  • Gemini 4 Argon — the first of the family, announced on 30 September 2026 and generally available from 13 October. Argon was Google’s answer to OpenAI’s GPT-6 Astra, built for long-horizon work with a 1M-token output limit. It was the frontier reasoning model aimed at cyber defenders first, then paid API users.
  • Gemini 4 Carbon — released 10 November 2026 as the flagship. Carbon is the model Google keeps for the most difficult long-horizon projects: deep mathematical and scientific problem solving, autonomous multi-step agents, and generation up to 1M tokens. It is the most capable — and, from 1 December, the most expensive.
  • Gemini 4 Hydrogen — released 12 November 2026 as the fast, everyday model. Hydrogen is the only Gemini 4 model with a free tier, and at $1.50/$7.50 it is the cheapest way into the family. It is built for speed and everyday coding and agent workflows.
  • Gemini 4 Neon — released 18 November 2026. Neon sits between Hydrogen and Carbon: deeper reasoning than Hydrogen, roughly twice the speed of Carbon, and — from December — cheaper per token than Carbon too.
Model Released API price (input / output per 1M) Free tier Best for
Gemini 4 Hydrogen 12 Nov 2026 $1.50 / $7.50 Yes Everyday coding and fast agent work
Gemini 4 Neon 18 Nov 2026 $3.00 / $15.00 No Frequent, complex coding and knowledge work
Gemini 4 Carbon 10 Nov 2026 $2.00 / $10.00 until 30 Nov, then $4.00 / $20.00 No Hardest long-horizon projects
Gemini 4 Argon 30 Sep 2026 $2.00 / $10.00 until 30 Nov, then $4.00 / $20.00 No Long-horizon reasoning and cyber defence

The comparison table hides one quirk worth knowing. Until 30 November, Carbon and Argon are actually cheaper per token than Neon, because they are still on their $2/$10 introductory rate. From 1 December that flips: Carbon and Argon rise to $4/$20, and Neon becomes 25% cheaper per token than either of them. If you are comparing models this week, compare on the December price, not the November one — unless your workload is genuinely a one-off.

There is also a quiet changing of the guard inside the Gemini app. Gemini 4 Neon is rolling out to Google AI Pro and Ultra subscribers, taking the Pro/Ultra slot previously held by Argon, and in Google Antigravity it takes the Pro orchestrator slot Argon used to occupy.

Argon has not been retired — it remains available through the API — but for subscribers, Neon is now the model you will reach for first. If you already read our Gemini 4 Argon guide or our Gemini 4 Hydrogen guide, think of Neon as the model that fills the gap those two left open: Argon’s depth without Argon’s price, Hydrogen’s speed without Hydrogen’s limits.

Gemini 4 Neon Benchmarks: The Numbers Google Published

Every benchmark number in this section comes from Google’s own launch post and the DeepMind model card, with independent leaderboard data from Artificial Analysis where noted. That distinction matters: vendor-reported scores are a reason to run your own test, not a purchasing decision. Independent evaluations of Gemini 4 Neon had not landed at the time of writing, and Google has published no benchmark table in the model card itself — the scores below come from the launch announcement.

Programmer using Gemini 4 Neon on dual monitors
Gemini 4 Neon’s strongest results are in coding: it leads the Artificial Analysis Coding Agent Index.

Gemini 4 Neon coding and agent benchmarks

Coding is where Gemini 4 Neon makes its boldest claim. Google reports that Neon scores 58.6 on the Artificial Analysis Coding Agent Index, ranking first overall — ahead of every other model Artificial Analysis has tested on agentic coding tasks. On the same index, Artificial Analysis measured Neon at xhigh thinking at 59, three points ahead of Carbon at high thinking (55), and in line with Claude Code with Fable 5.2 (58) and Claude Opus 5.5 (58), at a lower cost per task.

The supporting coding scores are close to Carbon across the board:

  • Terminal-Bench 4.1: 65.2% for Neon, against 65.8% for Carbon — a gap of just 0.6 points.
  • CursorBench 4.1: 65.4% for Neon, beating Carbon (58.1%) and Hydrogen (57.3%) by more than seven points. Cursor itself reports a slightly lower 63.7% for Neon at default effort, up from 57.3% for Hydrogen — the difference comes from different test settings, and both point the same way.
  • FrontierCode 2.0: 46.2% for Neon, against 47.3% for Carbon.
  • SWE-sweep: 39.0% inside Google’s Antigravity CLI harness, which Google calls state of the art — but see the harness warning below.

On agent and automation work, Neon also leads on several task-based benchmarks: first place on the Zapier AutomationBench at 56.3%, first on Legal Research Bench and Tax Agent Bench, and first on Vending-Bench 2 — a simulation in which a model runs a vending machine business over time — finishing with $16,843 at the end of the run.

The harness gap: one number that needs context

Here is the honest caveat. That 39.0% SWE-sweep score came from Google’s own Antigravity CLI harness. On a simpler, shared mini-SWE harness, the same model scored just 12.8%.

That is not a contradiction — different harnesses give the model different tools, different amounts of context, and different ways to retry — but it is a warning against quoting any single coding score as a fact about the model. The harness matters as much as the model, and Google’s own numbers show the gap can be enormous. Treat the 39.0% as “Neon performs well inside Google’s own tooling” and the 12.8% as “results will vary a lot depending on how you run it.” If you evaluate Gemini 4 Neon, do it inside the tools you actually use.

General intelligence and knowledge work

Outside coding, Gemini 4 Neon holds up well against the current frontier — again, on Google’s own comparison table, which pits it against Claude Opus 5.5 and OpenAI’s GPT-6.2 Sol:

  • Artificial Analysis Intelligence Index: 61 at xhigh thinking, ranking third overall — ahead of Claude Opus 5.5 (59) and GPT-6.2 Sol (57). Google says Neon leads on 16 of the 19 results in its comparison table.
  • Vals Index v2.2: 70.9%, placing second and edging Carbon’s 70.1%.
  • Speed: roughly 101–110 tokens per second, about twice Carbon’s ~57 tokens per second, according to third-party analysis of the launch data.

Neon is behind Opus 5.5 on Terminal-Bench 4.1 and FrontierCode 2.0, and behind GPT-6.2 Sol on Chartography, a chart-understanding benchmark. That pattern is consistent with what Google is selling: a model that is very strong at the practical work of coding agents and knowledge tasks, but not the outright leader on every deep-reasoning test. For a full comparison with the model it most directly challenges, see our GPT-6.2 Astra and Sol guide, and for the Claude side of the same fight, our Claude Fable 5.2 guide.

Gemini 4 Neon Pricing: The Full Rate Card

Gemini 4 Neon costs $3.00 per million input tokens and $15.00 per million output tokens on the standard API tier. That is the headline, but the full rate card has four tiers, and picking the right one can halve your bill. The figures below are taken directly from Google’s official Gemini API pricing page, checked on 19 November 2026.

Data centre servers powering Gemini 4 Neon
Every Gemini 4 Neon token is billed — including the thinking tokens the model spends before answering.
API tier Input (per 1M tokens) Output (per 1M tokens) Cached input (per 1M)
Standard $3.00 $15.00 $0.15
Batch $1.50 $7.50 $0.075
Flex $1.50 $7.50 $0.075
Priority $5.40 $27.00 $0.27

Three pricing facts are easy to miss. First, cached input costs $0.15 per million tokens — a 95% discount on the standard input price — but only implicit caching is supported; you cannot set up explicit cache entries for Neon. Second, Batch and Flex both cut the price by exactly half, to $1.50/$7.50, which makes them the obvious choice for anything that is not time-sensitive. Third, the Priority tier costs 80% more than Standard, so it is only worth it when latency genuinely matters.

There is also the question of how you pay as a consumer rather than a developer. Gemini 4 Neon is rolling out to Google AI Pro subscribers ($19.99 per month) and Google AI Ultra subscribers ($249.99 per month) in the Gemini app. For Pro subscribers, that is the real story of this launch: a frontier-class model at the cheaper subscription tier, without paying for Ultra. Free Gemini users do not get Neon at all, in the app or the API.

The December 1 twist that changes the maths

Neon’s $3/$15 rate looks odd next to Carbon’s $2/$10 — until you check the calendar. Carbon’s and Argon’s $2/$10 rate is an introductory price that ends on 30 November 2026. From 1 December, both rise to $4.00 input and $20.00 output per million tokens. At that point Gemini 4 Neon becomes 25% cheaper per token than Carbon, while running at roughly twice Carbon’s speed.

Google frames the comparison in cost per task rather than cost per token, and on that measure Neon looks better still. At its default high thinking level, Google says Neon costs 32% less per task than Carbon, and 50–80% less per task than the leading frontier models.

Artificial Analysis, measuring at xhigh thinking, puts the saving at about 17% per task against Carbon. The two figures differ because they measure at different thinking levels — xhigh spends more tokens per task — but both agree on the direction: Neon does the same work for meaningfully less money.

What a real task actually costs

Token prices are abstract, so here is what they mean in practice. A typical coding-agent session — say, 200,000 input tokens of repository context and 20,000 output tokens of code and explanation — costs about $0.90 on the standard tier ($0.60 input + $0.30 output). Run the same session on the Batch tier and it costs $0.45. Send the same repository context repeatedly with implicit caching, and the input side of the bill falls to almost nothing: 200,000 cached input tokens cost just $0.03.

Scale that up and the difference between tiers becomes real money. A small team running 100 such sessions a day spends roughly $90 a day on Standard, $45 on Batch, and a lot less again if most of the context is cached. That is why the practical advice for Gemini 4 Neon is simple: use Batch for anything that can wait, lean on caching for repeated context, and keep the thinking level at high unless a task genuinely needs xhigh. Thinking tokens are billed as output, so xhigh does not just think longer — it costs more for every extra token it spends thinking.

How to Use Gemini 4 Neon

Gemini 4 Neon is available in more places on day one than most Google launches manage, and the route you take depends on whether you are a subscriber, a developer, or an IDE user.

Developers collaborating with Gemini 4 Neon
Teams can reach Gemini 4 Neon through the Gemini app, the API, or directly inside their coding tools.

In the Gemini app

If you pay for Google AI Pro ($19.99/month) or Google AI Ultra ($249.99/month), open the Gemini app and pick Gemini 4 Neon from the model menu. The rollout is gradual, so if you do not see it yet, check again over the coming days — Google says it is reaching Pro and Ultra accounts in waves. In the app, Neon is best used for long documents, spreadsheets, PDFs, and charts in a single prompt: paste a research paper, a contract, or a set of quarterly reports and ask it to work across all of them at once. That is what the 1M-token window is for. Free users will not see the model at all.

In the API and AI Studio

Developers can call Gemini 4 Neon through the paid tier of the Gemini API with the model ID gemini-4-neon, or try it in Google AI Studio. You will need a paid API setup first — there is no free tier to experiment with, which is a real difference from Hydrogen. Android Studio also lists the model, and Gemini for business offers it through both the API and the model dropdown. If you are building an agent, the Batch tier is the sensible default: the same model at half price, for any job that does not need an instant answer.

In your coding tools

This is where Gemini 4 Neon may matter most, because Google shipped it into the tools developers already live in:

  • Google Antigravity: Neon is rolling out to paid Pro and Ultra users in the desktop app, CLI, and IDE extensions, taking the Pro orchestrator slot previously held by Argon.
  • GitHub Copilot: generally available in Copilot Pro+, Max, Business, and Enterprise — in VS Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, the Copilot CLI, the coding agent, github.com, and GitHub Mobile. Copilot Free and Pro are not on the list.
  • Cursor: available now, with Cursor reporting 63.7% on CursorBench 4.1 at default effort.
  • Vercel AI Gateway and OpenRouter: both listed the model on launch day, so it is reachable from any tool that speaks those gateways.

The Copilot and Cursor integrations are the ones to watch. A model that tops the Coding Agent Index is only useful if it is inside your editor, and on day one Neon is already there. If you use either tool, switching the model dropdown is a five-second test — run your usual tasks for a day and compare the results against whatever you were using before, rather than trusting a benchmark table.

Picking a Gemini 4 Neon thinking level

Gemini 4 Neon ships with four thinking levels — low, medium, high, and xhigh — and high is the default. Think of them as a dial between speed and depth:

  • Low is for quick, simple tasks: short answers, simple code edits, fast classification. It spends the fewest tokens and answers fastest.
  • Medium suits everyday coding and drafting, where you want a bit of reasoning without paying for deep deliberation.
  • High is the default and the right choice for most complex work: multi-file refactors, long documents, and agent tasks with several steps.
  • Xhigh adds deeper multi-step reasoning for the hardest problems — but it spends more tokens, costs more, and third-party reports note it can make the model’s chain of thought ramble on some tasks.

The practical rule is to start at high, drop to medium or low for routine work, and only reach for xhigh when a task has genuinely defeated the default. Every thinking token is billed as output, so the dial is also a spending control.

Gemini 4 Neon vs GPT-6.2 Sol vs Claude Opus 5.5

Google’s launch post compares Gemini 4 Neon directly with the two models most buyers will weigh it against: OpenAI’s GPT-6.2 Sol and Anthropic’s Claude Opus 5.5. Here is how that comparison looks, using Google’s own numbers — so read it as Google’s case for Neon, not as an independent verdict.

Measure Gemini 4 Neon GPT-6.2 Sol Claude Opus 5.5
Artificial Analysis Intelligence Index 61 (3rd overall) 57 59
Artificial Analysis Coding Agent Index 58.6 (1st overall) Not reported 58 (via Claude Code + Fable 5.2)
API price (input / output per 1M) $3.00 / $15.00 $2.00 / $10.00 Not in Google’s table
Context window 1M tokens Not in Google’s table Not in Google’s table
Cost per task vs Carbon 32% lower (default high) — —

On price, GPT-6.2 Sol is actually cheaper per token than Neon — $2/$10 against $3/$15 — so Neon’s argument is not that it is the cheapest model on the market. Its argument is that it is the cheapest way to get this level of coding-agent performance, at roughly twice the speed of Google’s own flagship, with a 1M-token window that lets it hold an entire repository in context. Against Claude Opus 5.5, Neon’s edge is the benchmark table: ahead on 16 of 19 results Google lists, including the Intelligence Index, while Opus 5.5 still leads on Terminal-Bench 4.1 and FrontierCode 2.0. Against GPT-6.2 Sol, Neon leads on the Intelligence Index but trails on Chartography.

There is a wider context too. OpenAI’s GPT-6.2 Astra and Sol launched on 17 November, the day before Neon — so this is a direct, same-week fight for the same buyers.

And Meta’s Muse Ember, which we covered earlier this week, is the budget challenger in the same agent-workload space at $0.10/$0.50 per million tokens. The market is splitting into three lanes: tiny cheap models for routing and extraction, mid-tier workhorses like Neon for daily complex work, and flagships for the hardest problems. If you want the full picture of the budget lane, see our Meta Muse Ember guide.

Who Should Use Gemini 4 Neon — and Who Should Not

Team reviewing work built with Gemini 4 Neon
For teams already paying for Google AI Pro, Gemini 4 Neon is the most capable model on the cheaper subscription tier.

Gemini 4 Neon is a strong fit if you:

  • Already pay for Google AI Pro. Neon is the most capable model available at the $19.99/month tier, and you get it without upgrading to Ultra. This is the single best reason to try it first.
  • Run coding agents all day. Repository-scale refactors, multi-file edits, and tool-heavy agent loops are exactly what Neon was tuned for, and its number-one Coding Agent Index ranking is the most meaningful score Google published.
  • Work with long documents. Contracts, research papers, financial reports, and legal research across spreadsheets and PDFs in one prompt — Neon’s 1M-token window and its first-place finishes on Legal Research Bench and Tax Agent Bench point at this use case.
  • Build agents on a budget. At $1.50/$7.50 on Batch, with cached input at $0.15, Neon is a genuinely cheap way to run a capable agent in production — provided your jobs can wait for batch processing.
  • Use GitHub Copilot or Cursor. Neon is already in both, so testing it costs you a dropdown change, not a migration.

Gemini 4 Neon is probably not the right pick if you:

  • Want a free model. There is no free tier. Hydrogen is free in the API and in the Gemini app, and it is the right starting point if you are not ready to pay.
  • Need the deepest reasoning available. For the hardest long-horizon problems — deep science, complex mathematics, autonomous agents running for hours — Carbon is still Google’s flagship, and its $2/$10 introductory price is actually cheaper than Neon until 30 November.
  • Need image or audio generation, or live voice. Neon outputs text only and does not support the Live API. Look elsewhere for multimodal generation.
  • Need the freshest possible knowledge. The March 2026 cutoff means anything from the last eight months is outside its training data, and Google warns some domains may be less current.

Gemini 4 Neon Limitations and Honest Caveats

A launch-day article that only repeats the press release is not much use, so here are the limits that matter — most of them from Google’s own documents.

The benchmarks are Google’s own. The scores in this article come from Google’s launch post and the DeepMind model card, not from independent re-testing. Artificial Analysis has published its own leaderboard numbers for Neon, and those broadly agree with Google’s framing, but independent coding evaluations had not landed at the time of writing. Treat every score as a claim to verify, not a fact.

The safety report is not finished. The model card marks the full Safety and Security Report as “coming soon”.

Google carried forward its Frontier Safety Framework results from Gemini 4 Argon, because Neon’s capability gain over Argon stayed below the threshold that would trigger a full new evaluation — but that also means the carried-forward findings include unresolved risk thresholds in biosecurity and cybersecurity. A separate cyber variant, “Neon for cyber”, is rolling out to vetted members of Google’s Fairwind Program over the days following launch, which is a reminder that the frontier version of this technology is deliberately restricted.

Hallucinations remain possible. The model card says so directly. A 1M-token window does not stop a model from confidently inventing a citation, a function, or a clause in a contract. For legal and financial work — the very areas where Neon scores well — human review is not optional.

Thinking tokens cost real money. Every token Neon spends thinking is billed as output at $15 per million. At xhigh thinking on a long task, the thinking can cost more than the answer. If you leave the dial at xhigh by default, your bill will show it.

The December price rise cuts both ways. Neon’s advantage over Carbon only appears on 1 December. Until then, Carbon is cheaper per token. And Neon’s own price is not locked in forever — Google has already shown it will move these prices, and the November 30 expiry on Carbon’s intro rate is proof of that. Check the official pricing page before you commit a budget to any of these numbers.

Gemini 4 Neon FAQs

Is Gemini 4 Neon free?
No. Gemini 4 Neon has no free tier in the API or in the Gemini app. Developers need the paid tier of the Gemini API, and consumers need a Google AI Pro ($19.99/month) or Ultra ($249.99/month) subscription. If you want a free Gemini 4 model, Gemini 4 Hydrogen is the one.

How much does Gemini 4 Neon cost?
On the standard API tier, $3.00 per million input tokens and $15.00 per million output tokens, with cached input at $0.15. Batch and Flex tiers halve that to $1.50/$7.50, and Priority costs $5.40/$27.00. Thinking tokens are billed as output.

What is the Gemini 4 Neon context window?
1,048,576 tokens — roughly one million — shared across input, thinking, and output. A normal response can run up to 65,536 tokens, and a background long-decode mode can stretch output up to the full 1M-token budget.

Is Gemini 4 Neon better than GPT-6.2 Sol?
On Google’s own comparison table, Neon scores 61 on the Artificial Analysis Intelligence Index against 57 for GPT-6.2 Sol, and leads on 16 of 19 listed results. GPT-6.2 Sol is cheaper per token at $2/$10, and it leads on Chartography. Independent tests have not yet settled the question, so the honest answer is: Neon looks stronger on coding-agent work, Sol is cheaper, and you should test both on your own tasks.

Where can I use Gemini 4 Neon?
In the Gemini app (Pro and Ultra subscribers), the Gemini API and Google AI Studio (paid tier), Android Studio, Gemini for business, Google Antigravity, GitHub Copilot (Pro+, Max, Business, Enterprise), Cursor, Vercel AI Gateway, and OpenRouter.

What are Gemini 4 Neon’s thinking levels?
Four levels — low, medium, high, and xhigh — that trade answer depth against speed and token spend. High is the default. Use low or medium for routine work, high for complex tasks, and xhigh only for the hardest problems, since thinking tokens are billed.

Does Gemini 4 Neon replace Gemini 4 Argon?
Not exactly. Argon remains available through the API, but Neon has taken over Argon’s slot for Pro and Ultra subscribers in the Gemini app and in Antigravity’s Pro orchestrator. For most subscribers, Neon is now the model you will use day to day.

Conclusion

Gemini 4 Neon is the most practical model Google has shipped in the Gemini 4 family so far — not the most powerful, not the cheapest, but the one most people will actually use. It takes Carbon’s depth, trims it into a faster and cheaper package, tops the Coding Agent Index, and lands inside the tools developers already open every morning.

The pricing twist is real but temporary: until 30 November, Carbon is cheaper per token on its introductory rate; from 1 December, Neon is 25% cheaper and twice as fast, which is the comparison that will matter for the next year.

The sensible way to judge it is the boring way: try it in your own editor for a day. Switch the model in Cursor or Copilot, run the tasks you run every day, and compare the results and the bill against whatever you used before. The benchmark table says Gemini 4 Neon should win that test for most coding and knowledge work — but the harness gap in Google’s own numbers is a reminder that the only score that matters is yours.

Leave a Comment

Your email address will not be published. Required fields are marked *