Google shipped its fourth model in about seven weeks, and I had to make a chart just to keep the names straight. Hydrogen, Carbon, Argon, and now Gemini 4 Neon, which landed on November 18, 2026.

So I spent launch day doing the boring part for you. I read Google's announcement, the DeepMind model card, the AI Studio migration guide, and the pricing page, then checked the independent numbers from Artificial Analysis.

Here's the straight version: what Neon costs, how good it actually is (including where its rivals still beat it), and when to pick it over Carbon or Hydrogen. No spreadsheet required.

Gemini 4 Neon at a glance

DetailGemini 4 Neon
ReleasedNovember 18, 2026
Model IDgemini-4-neon (stable)
Where it sitsBetween Hydrogen (everyday) and Carbon (hardest work)
Price per 1M tokens$3 input / $0.15 cached input / $15 output
Context window1,048,576 tokens (shared input, thinking, output)
Max output65,536 synchronous, up to 1M with long decode
Inputs / outputsText, image, video, audio, PDF in; text out
Thinking levelsLow, medium, high (default), xhigh
Knowledge cutoffMarch 2026
TierPaid only (no free tier)

What is Gemini 4 Neon?

Gemini 4 Neon is Google's new mid-tier model, built for complex coding, knowledge work, and science that you run often enough to care about speed and cost. Google's pitch is "frontier-level reasoning for complex, high-frequency workflows, faster and more cost efficiently."

The model card is refreshingly plain about its origins: Neon is adapted from Gemini 4 Carbon. So you get much of Carbon's quality, tuned to answer quicker and bill less.

Where Neon fits in the Gemini 4 family

Google now names its models after elements, and the lineup maps to a real hierarchy. Hydrogen is the fast everyday worker. Carbon is for your hardest, long-horizon projects. Neon slots into the busy middle: complex work you do all the time. 9to5Google describes Hydrogen as the replacement for the Flash line, and our Gemini 3.8 Flash guide covers the older fast model if you still run it.

If you want the blow-by-blow, we put together a full Gemini 4 Neon vs Gemini 4 Carbon breakdown. The short version is that Neon trades a little peak capability for a lot more speed. One Google launch partner, Glean, said Neon matched or beat Carbon on quality in their early evaluations at roughly three times the speed and 60% of the cost. That's a vendor-friendly data point, so treat it as a claim and lean on the independent numbers below.

There's also a shuffle in the consumer app. 9to5Google reports that Neon replaces Gemini 4 Argon in the Gemini app picker (showing up as "Deeper analysis"), and that the older 3.1 Pro is being retired. Google's own pages confirm Neon is in the app for Pro and Ultra subscribers, so I'd trust the availability and treat the "replaces Argon" framing as 9to5Google's read until Google says it outright.

And yes, keeping Argon, Hydrogen, Carbon, and Neon apart is genuinely confusing. You're not imagining it.

Explore integrations

Gemini 4 Neon pricing

Gemini 4 Neon pricing is flat and easy to reason about: $3 per million input tokens, $0.15 per million cached input tokens, and $15 per million output tokens, the same rate across the whole 1M-token window. It's on the paid tier only, so there's nothing to try on the free tier.

Output tokens include thinking tokens, so the model's internal reasoning bills at the $15 output rate. At the higher thinking levels, that reasoning adds up fast.

Here's how Neon lines up against its siblings. For the full picture across the lineup, see our Gemini pricing guide.

💲 Model📥 Input / 📤 output per 1M📝 Note
Hydrogen$1.50 / $7.50Neon costs about 2x Hydrogen
Neon$3 / $15Flat across the 1M window
Carbon and Argon$2 / $10 then $4 / $20$2 / $10 through Nov 30, 2026, then $4 / $20 from Dec 1

That schedule creates a quirk worth planning around. Until November 30, Carbon is cheaper per token than Neon. From December 1, Carbon's standard price jumps and Neon becomes about 25% cheaper. So the "cheaper option" between the two literally flips based on your calendar.

A quick worked example at list rates. Say a task sends 50,000 input tokens and generates 20,000 output tokens (thinking included). The input costs about $0.15, the output about $0.30, so roughly $0.45 for the task. If most of that input hits the implicit cache, the input side drops to under a cent. Multiply by a few thousand runs and the thinking-token behavior is the thing that moves your bill.

A few more pricing details from Google's pricing page: the Batch API runs at a 50% discount, and Flex and Priority inference are both supported. Grounding with Google Search gives you 5,000 free requests a month shared across Gemini 3 and newer models, then charges $14 per 1,000 requests.

Google's official benchmarks (the vendor view)

These numbers come from Google's own blog and model card, so read them as the vendor's scorecard. They look strong, and I'll point out where they aren't.

Google reports Neon at 61 on the Artificial Analysis Intelligence Index (version 5.0.1), ahead of Claude Opus 5.5 at 59 and GPT-6.2 Sol at 57. It claims the top spot on Zapier's AutomationBench (56.3%) and on Andon Labs' Vending-Bench 2, where it ran a simulated business to $16,843. It also debuts at #2 on Arena's Agent Arena Index with a score of 1336.

🧪 Benchmark🔵 Neon🟣 Opus 5.5🟢 GPT-6.2 Sol
AA Index v5.0.1615957
SWE-sweep (provider harness)39.0%10.1%8.3%
AutomationBench56.3%42.5%41.3%
Terminal-Bench 4.1 (cheating-adjusted)65.2%65.8%61.5%
FrontierCode 2.046.2%47.3%43.8%
Chartography73.5%66.3%75.1%
HLE-Diamond71.3%67.9%64.0%

Neon doesn't sweep. Claude Opus 5.5 edges it on Terminal-Bench 4.1 (65.8% to 65.2%) and on FrontierCode 2.0 (47.3% to 46.2%), and GPT-6.2 Sol tops it on Chartography for reading dense charts (75.1% to 73.5%). Close races, but Google's own table shows the losses, which is more than most launch posts admit.

What independent testing shows

Artificial Analysis published its own launch-day numbers, and this is the part I'd anchor on, because they test everyone the same way.

On the Artificial Analysis Intelligence Index, Neon scores 61 at xhigh, 60 at high, 55 at medium, and 44 at low. Per Artificial Analysis's launch article, that 61 is the third-highest score overall, behind GPT-6.2 Astra and Claude Fable 5.2, both running at their max settings. One point ahead of Carbon, so this is a real but small step up.

The efficiency story is the headline. Artificial Analysis measured Neon at about $6.44 per Index task against Carbon's $7.76, and roughly 21 minutes per task against Carbon's 36. Neon pushes about 110 output tokens per second versus Carbon's 57. The catch, and there's always a catch: Neon writes about 15% more output tokens per task, so it's chatty, especially at xhigh.

Coding agents tell a similar story. Using the Antigravity CLI, Artificial Analysis put Neon (xhigh) at 59 on its Coding Agent Index, ahead of Carbon (high) at 55 and in line with Claude Fable 5.2 and Claude Opus 5.5 at 58, both tested in Claude Code. The difference is cost: about $6.86 per task for Neon versus $12.61 for Opus 5.5 and $15.47 for Fable 5.2. Same ballpark on quality, a fraction of the bill.

Context window, thinking levels, and API notes

Neon keeps the Gemini 4 context window: a unified 1,048,576 tokens shared across input, thinking, and output. Synchronous responses cap at 65,536 tokens, and you can push past that up to the full 1M using background execution (long decode).

It takes text, images, video, audio, and PDFs as input and returns text. Live API, audio generation, and image generation aren't supported here, so Neon reads and writes but won't produce audio or images for you.

The Gemini 4 Neon API behaves like the rest of the family, with a few changes worth knowing before you migrate:

  • Thinking levels: low, medium, high (default), and xhigh. Neon rejects minimal, and it defaults to high where Carbon and Argon defaulted to medium.
  • Caching: explicit caching is gone. Implicit caching kicks in automatically on prompts of at least 4,096 tokens and discounts cached input by 95%.
  • Sampling: temperature, top_p, and top_k are ignored. Send out-of-range values and you'll get a 400 error, so strip them.
  • Tools and state: declare tools natively and preserve thought signatures across turns.
  • Timeouts: hard prompts can run for minutes, so set client timeouts to at least 15 minutes.

The knowledge cutoff is March 2026, so for anything newer you'll want grounding or your own context.

Limitations worth knowing

No model is all upside, and Neon has a clear list of tradeoffs:

  • Paid tier only. Nothing to test on the free tier.
  • No Live API, audio generation, or image generation. Text output only.
  • Verbose at xhigh. The extra output tokens cost real money at the $15 rate.
  • Hard prompts are slow. Think minutes on the toughest work, sometimes several.
  • Explicit caching removed and sampling params ignored. Small code changes if you're migrating.
  • March 2026 knowledge cutoff. Recent events need grounding.

On safety, Google took a shortcut that it's open about. Because Neon's capabilities sit close to Argon's, Google reused Argon's Frontier Safety determinations and didn't rerun the full suite. Three of those determinations are notable: the biological CBRN uplift and cybersecurity uplift critical capability levels "cannot be ruled out," and neither can its "loss of control" threshold for stealth and situational awareness. Google says mitigations apply in each case. The full Safety and Security Report is still "coming soon," and Google is rolling out Neon's cyber capabilities to trusted defenders through its Fairwind Program.

Should you use Neon, Carbon, or Hydrogen?

Most of this comes down to how hard and how frequent your work is.

Choose Neon if: you run complex coding, analysis, or science often, you want close-to-Carbon quality with more speed, and you need the full 1M-token window. It's the default for heavy day-to-day work.

Choose Carbon if: you're tackling the hardest, long-horizon problems where peak capability wins and cost is secondary (our Gemini 4 Carbon vs Claude Opus 5.5 comparison covers that top end). Remember the pricing flip: from December 1, Neon actually costs less per token than Carbon, so "use Carbon to save money" stops being true.

Choose Hydrogen if: you want fast, cheap everyday execution and don't need frontier reasoning on every call. In Google Antigravity, you can even pair them, with Neon planning and Hydrogen executing the routine steps.

As for Argon, it's effectively the one Neon steps in front of (we covered Argon on its own in Gemini 4 Argon vs GPT-6 Astra). 9to5Google reports Argon is being replaced in the Gemini app, and Neon is the newer middle option you'd reach for instead. If you're weighing the lineup, our guide to choosing a Gemini model walks through it in more depth.

Where an AI employee fits once you've picked a model

Picking a model is the fun part, but a model on its own still waits for you to prompt it, so turning "I have access to a frontier model" into finished work is still on you.

That's the job Lindy does. Lindy is an AI employee you hand real work to, by text or in your team's channels, and it's model-agnostic: you pick the model for each task and focus on the outcome while it handles the steps. Which model wins the week, Gemini 4 Neon included, stops being your problem.

With Lindy, teams can:

  • Triage and route email: route the right messages to the right place.
  • Schedule meetings and take notes: book calls and capture the follow-ups automatically.
  • Keep your CRM updated: log activity and update records after calls and emails.
  • Run quick research and summaries: pull together the context you need before a decision.
  • Handle follow-ups with an approval step: Lindy checks with you before anything sends.

The model race will keep moving, and next week there'll be another element on the periodic table. An AI employee that works across whatever's best lets you stop chasing it.

Want to see it on your own work? Try Lindy free and hand off a task today.

Frequently asked questions

What is Gemini 4 Neon?

Gemini 4 Neon is Google's mid-tier Gemini 4 model, released November 18, 2026, and adapted from Gemini 4 Carbon. It's built for complex coding, knowledge work, and science, and sits between Hydrogen and Carbon in the lineup.

How much does Gemini 4 Neon cost?

Gemini 4 Neon costs $3 per million input tokens, $0.15 per million cached input tokens, and $15 per million output tokens, flat across its full context window. It's available on the paid tier only, and thinking tokens bill at the output rate.

What is Gemini 4 Neon's context window?

Gemini 4 Neon has a unified context window of 1,048,576 tokens, shared across input, thinking, and output. Synchronous responses cap at 65,536 output tokens, and long decode can extend output up to the full 1M.

Is Gemini 4 Neon free?

No, Gemini 4 Neon is not free. It runs on the paid tier of the Gemini API only, though paid Google AI Pro and Ultra subscribers can also use it in the Gemini app and Google Antigravity.

Where can I use Gemini 4 Neon?

You can use Gemini 4 Neon 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 model picker for Pro and Ultra users.

Gemini 4 Neon vs Carbon: what's the difference?

The main difference between Neon and Carbon is speed versus peak capability. Carbon aims for the hardest, long-horizon tasks, while Neon gives you close-to-Carbon quality with more speed and, from December 1, 2026, a lower price per token. Our full Gemini 4 Neon vs Gemini 4 Carbon breakdown goes test by test.

Gemini 4 Neon vs Argon: what's the difference?

The main difference between Neon and Argon is that Neon is the newer middle-tier model built to cover high-frequency complex work. Argon is the older middle option, and per 9to5Google it's the one Neon replaces in the Gemini app picker.

Does Gemini 4 Neon replace Argon?

Per 9to5Google, yes, Neon replaces Argon in the Gemini app model picker, where it appears as "Deeper analysis." Google's own pages confirm Neon's availability but don't frame it as a direct replacement, so treat that detail as reported by 9to5Google.

What is Gemini 4 Neon's knowledge cutoff?

Gemini 4 Neon has a knowledge cutoff of March 2026. For events or information after that date, you'll need to supply context or use grounding with Google Search.