Google now ships two top-tier Gemini 4 models, so the Gemini 4 Neon vs Carbon question is really Google versus Google. Neon landed on November 18, 2026, adapted from Carbon, and I read through Google's launch post, both model cards, the Gemini API pricing page, and Artificial Analysis's numbers to sort out which one you should actually run. By the end you'll know which one to run, and what it costs at both Carbon price points.

Gemini 4 Neon vs Carbon: the short answer

Neon matches Carbon on raw intelligence and beats it on speed and cost. On Artificial Analysis's Intelligence Index, Neon scores 61 at its xhigh thinking level against Carbon's 60 at high. That's a slim one-point gain, so treat the two as roughly even on raw smarts.

The difference shows up everywhere else. Neon runs at roughly 110 tokens per second to Carbon's 55, finishes an Index task in about 21.5 minutes against Carbon's 37, and costs less per task. Google's own launch post puts Neon at a 32% lower cost per task than Carbon at the high thinking level.

So the two models split cleanly by job.

Choose Neon if you:

  • Want frontier-level results with fast, cheap end-to-end task completion.
  • Run high-frequency, complex coding, knowledge, or science work.
  • Want four thinking levels, including xhigh for the occasional hard problem.
  • Plan to keep using it past December 1, when it becomes the cheaper model per token.

Choose Carbon if you:

  • Take on the hardest architecture, debugging, science, and security problems.
  • Need the specific benchmarks where Carbon still leads, like Agents' Last Exam and very long-context reasoning.
  • Are buying tokens before November 30, while Carbon is cheaper per token.
  • Want Carbon's wider distribution through Google AI Mode and the Fairwind Program.

Consider Hydrogen if: you mostly want everyday speed. In a Gemini 4 Neon vs Hydrogen choice, Hydrogen is the everyday lane at half Neon's token price. The older Gemini 4 Argon costs the same as Carbon on the API, and in the Gemini app 9to5Google reports that Neon has taken its place in the model picker, so there is little reason to start on it now.

Meet Gemini 4 Neon

Gemini 4 Neon is Google's new model for complex coding, analytical reasoning, and production-grade enterprise work. Google's framing is that it delivers frontier reasoning "faster and more cost efficiently," and the model card says plainly that Neon is adapted from Carbon. Our Gemini 4 Neon explainer has the full launch details.

In the Gemini 4 family, Neon sits between Hydrogen, which handles fast everyday execution, and Carbon, which takes the hardest long-horizon problems. Neon is the middle gear built for high-frequency complex work.

It ships with a unified 1 million-token context window and four thinking levels: low, medium, high (the default), and xhigh. At xhigh it ranks third overall on Artificial Analysis's Intelligence Index with a score of 61, behind only larger, slower models.

Meet Gemini 4 Carbon

Gemini 4 Carbon is Google's most capable reasoning model, built for jobs that take a long time. Google says teams have watched it "run for days," working through dead ends to root-cause and patch subtle bugs across large codebases. It is based on Gemini 4 Argon, and Google announced it a few weeks before Neon. Our full write-up on who can use Carbon and what it costs has the deeper background.

Carbon's sweet spot is the hardest architecture, debugging, science, and security problems. It defaults to the medium thinking level and caps at high, so it has fewer gears than Neon but a very high ceiling.

Carbon also reaches more places. On top of the API and the Gemini app, it is listed in Google AI Mode and the Fairwind Program for cyber defenders, and its consumer access runs through Google AI Ultra.

Explore integrations

Gemini 4 Neon vs Carbon: specs and pricing at a glance

Here is the side-by-side. The pricing rows are where most of the real difference lives, so read those twice.

🔧 Spec🟦 Gemini 4 Neon🟨 Gemini 4 Carbon
ReleasedNovember 18, 2026November 2026 (announced Oct 28)
Best forFast, complex, high-frequency workThe hardest long-horizon problems
Input price (per 1M)$3.00 (flat)$2.00 through Nov 30, then $4.00 from Dec 1
Cached input (per 1M)$0.15$0.10, then $0.20 from Dec 1
Output price (per 1M)$15.00 (flat)$10.00 through Nov 30, then $20.00 from Dec 1
Context window1,048,576 tokens1M tokens
InputsText, images, audio, video, PDFText, images, audio, video, PDF
OutputText (64K default, up to 1M via long decode)Text (64K default, up to 1M via long decode)
Thinking levelslow, medium, high, xhighmedium (default), caps at high
CachingImplicit only (95% off)Implicit only (95% off)
Knowledge cutoffMarch 2026March 2026
API tierPaid onlyPaid only

Both read spreadsheets, PDFs, and photos in the Gemini app, both share one 1M-token window across input, thinking, and output, and neither supports explicit caching. So let's run the pricing numbers.

Gemini 4 Neon cost vs Carbon: what each model costs

Gemini 4 Carbon vs Neon pricing tells two different stories depending on the date.

Neon is flat: $3 per million input tokens, $0.15 cached, and $15 per million output tokens, with no scheduled change. Carbon is on a timer: $2 input, $0.10 cached, and $10 output through November 30, 2026, then $4, $0.20, and $20 starting December 1. Our Gemini pricing guide tracks the full plan lineup.

So Carbon is cheaper per token until November 30. From December 1, Neon is 25% cheaper per token than Carbon.

Say your app sends 1,000 requests, each with 5,000 input tokens and 2,000 output tokens. Here is the bill three ways:

💵 Model and price📥 Input cost📤 Output cost🧾 Total
Neon (flat)$15$30$45
Carbon, through Nov 30$10$20$30
Carbon, from Dec 1$20$40$60

Through November, Carbon wins that workload. In December, Neon wins it by the same margin.

Cost per task tells a cleaner story, because it folds in how many tokens each model burns. Artificial Analysis puts Neon at $6.44 per Intelligence Index task against Carbon's $7.76 at standard pricing, a 17% gap on a setup that runs Neon at xhigh, smaller than the 32% Google claims at the high level. Neon is a bit more verbose, but it still lands cheaper per finished task.

If you are only pricing tokens, check the calendar. If you are pricing finished work, Neon already costs less.

Official benchmarks: Google's numbers

One trap first. The two model cards use different benchmark versions, so you cannot line every row up. Neon's card reports the Artificial Analysis Index v5.0.1 and Terminal-Bench 4.1; Carbon's card reports Index v5.0 and Terminal-Bench 4.0. I only compared rows that use the same version, and you should too.

Here are the directly comparable rows from Google's own tables. Treat all of these as vendor claims.

📊 Benchmark🟦 Neon🟨 Carbon🏆 Leader
Vals Index v2.270.9%70.4%Neon
Agents' Last Exam v1.175.8%77.4%Carbon
AutomationBench56.3%53.9%Neon
RiemannBench84.0%80.0%Neon
Terminal-Bench-Science 0.170.5%64.8%Neon
Chartography73.5%72.5%Neon
GraphWalks 256k to 1M95.2%96.2%Carbon
SEC-bench Pro90.5%88.2%Neon
Vending-Bench 2$16,843$16,484Neon

Neon leads most of the comparable rows, which fits a newer model adapted from Carbon. Carbon holds onto Agents' Last Exam and the long-context GraphWalks test, both the kind of deep, sprawling work it was built for. Google also lists Neon at number one on the Artificial Analysis Coding Agent Index with a 58.6.

Independent benchmarks: Artificial Analysis

Vendor tables tend to flatter the vendor, so I checked Artificial Analysis, which runs every model through the same setup. I pulled these on November 19, and AA's numbers move, so re-check before you quote them.

On the Intelligence Index, Neon scores 61 at xhigh and Carbon scores 60 at high. At the top of the ladder, this is close to a tie on raw intelligence.

The speed and latency numbers are where they separate:

  • Output speed: about 110 tokens per second for Neon, about 55 for Carbon.
  • Time to first answer token: 56.51 seconds for Neon at xhigh, 73.19 seconds for Carbon at high.
  • Time per Index task: about 21.5 minutes for Neon, about 37 for Carbon.
  • Cost per Index task: $6.44 for Neon, $7.76 for Carbon.

Both models are verbose, so your bills will skew toward output tokens either way. Neon generates a bit more per task than Carbon, which is the one place its efficiency slips.

Access and limits

Both models are paid-tier only on the API, so neither runs on the free tier. Every token counts, including thinking tokens, which bill as output.

Where Neon runs: the Gemini API, Google AI Studio, Gemini for business, Google Antigravity for paid AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra subscribers. Where Carbon adds reach: Google AI Mode and the Fairwind Program for cyber defenders, with consumer access through Google AI Ultra.

A few limits worth knowing before you build:

  • Caching is implicit only. Cached input is 95% off on prompts of at least 4,096 tokens.
  • Hard prompts can take minutes. Google recommends client timeouts of at least 15 minutes, and anything past 64K output tokens needs background execution for long decode.
  • Neon's safety paperwork is not done. Its Safety and Security Report is listed as "coming soon," and its Frontier Safety results are carried forward from Argon. Google states that the biological-uplift and cybersecurity-uplift critical capability levels cannot be ruled out, and neither can its loss-of-control threshold for stealth and situational awareness, with mitigations in place. For regulated or high-stakes work, that is worth a note in your own review.

Which should you use

Here is the Gemini 4 Neon vs Carbon decision in one place.

Pick Neon for high-frequency, complex coding, knowledge, and science work where speed and cost decide the day. Pick Carbon for the hardest, longest-running problems and the handful of benchmarks it still tops, like Agents' Last Exam and GraphWalks. For a cross-vendor gut check, our Gemini 4 Carbon vs Claude Opus 5.5 breakdown shows how Carbon fares outside the family, and our Gemini 4 Argon vs GPT-6 Astra piece does the same for Argon.

Then factor the calendar. On the 1,000-request example above, Carbon costs $30 against Neon's $45 through November 30, then $60 against the same $45 from December 1. If you are planning past the new year, price your real workload at both Carbon rates before you commit.

If you want to be sure, test on your own work:

  1. Pick 10 real tasks you do every week.
  2. Run both models at their default thinking level, then again at the top level, and note the cost, time, and quality.
  3. Price your results at both Carbon rates, so December's bill does not surprise you.

The catch is that the leaderboard changes most weeks, so whatever you pick today may not be the winner next month. That is easier to live with when you are not hard-wired to one model.

How to use the right model without betting on one

Model launches now come every couple of weeks. Neon landed three weeks after Carbon was announced, Argon shipped before both, and Claude and GPT keep trading the top spot too. Picking one model and hard-coding it into your stack is how you end up rebuilding every time the frontier moves.

Lindy is an AI employee you can text to handle daily work, and it is model agnostic, so you choose the model for each task and switch whenever a better one ships.

With Lindy, teams can:

  • Pick the model for each task, and switch when a better one ships.
  • Text Lindy in Slack, email, or SMS to get answers and get work done in plain English.
  • Triage, label, and route incoming email automatically.
  • Schedule meetings and record them, with notes and follow-up drafts ready for you to approve.
  • Update the CRM after calls and keep deal rooms current.
  • Build an employee knowledge base Lindy answers from across Slack, email, and SMS.
  • Hand off repetitive daily tasks, with Lindy asking before it sends anything.

So when the next model ships and the benchmarks reshuffle, your AI employee keeps working and you route the next job to whatever is winning. Try Lindy free and put it to work in a couple of minutes.

Frequently asked questions

Is Gemini 4 Neon better than Gemini 4 Carbon?

Gemini 4 Neon and Carbon are roughly tied on overall intelligence, scoring 61 and 60 on Artificial Analysis's Intelligence Index. Neon is faster and cheaper per task, while Carbon still leads a few of the hardest benchmarks and is Google's pick for the most demanding long-horizon work.

Which is cheaper, Gemini 4 Neon or Carbon?

Carbon is cheaper per token through November 30, 2026, at $2 input and $10 output. From December 1, Carbon rises to $4 and $20, which makes Neon ($3 and $15) about 25% cheaper per token. Per finished task, Artificial Analysis already puts Neon lower, at $6.44 against $7.76.

Does Gemini 4 Neon replace Gemini 4 Argon?

In the Gemini app, yes. 9to5Google reports that Neon replaces Argon as the "Deeper analysis" option in the model picker. In the Gemini API, the Gemini 4 Neon vs Argon choice is simpler: Argon is still listed, with the same pricing as Carbon, while Neon is the newer model adapted from Carbon for fast, high-frequency complex work.

What are the context windows for Neon and Carbon?

Both Gemini 4 Neon and Carbon have a unified 1 million-token context window, with Neon listed at 1,048,576 tokens. That single window is shared across input, thinking, and output tokens, and both support long decode up to 1M tokens.

Which is faster, Gemini 4 Neon or Carbon?

Gemini 4 Neon is faster. Artificial Analysis clocks it at about 110 tokens per second against Carbon's 55, with a quicker first token (56.51 seconds versus 73.19) and roughly 40% less time per Intelligence Index task.

Should I switch from Carbon to Neon?

Switch to Neon for high-frequency complex work where speed and cost matter most, and stay on Carbon for the hardest long-horizon problems and the benchmarks it still leads. Re-price both models after December 1, when Carbon's introductory pricing ends.