Categories AI and Tools

Gemini 4 Neon Is Live: $3/$15, Beats Carbon on Speed

Google launched Gemini 4 Neon on November 18, and the headline is not the model itself — it is the calendar. Neon costs $3 per million input tokens and $15 per million output tokens, flat, with no introductory price and no end date. Its bigger sibling, Gemini 4 Carbon, costs $2 and $10 until November 30, then doubles to $4 and $20 on December 1. That means for the next ten days the flagship is cheaper than the mid-tier model, and on December 1 the price ladder snaps into the shape Google actually designed: Hydrogen at $1.50/$7.50, Neon at $3/$15, Carbon at $4/$20.

If you are choosing a model for a system you will still be running in January, choose on the December prices, not today’s. This article gives you the full family price table, the independent benchmark numbers (kept separate from Google’s own claims), a worked bill for a real workload, and a straight answer on which model to pick for which job.

What Gemini 4 Neon actually is

Neon is the fourth Gemini 4 model in seven weeks, and it fills the gap Google left between its fast model and its flagship. Argon arrived October 13, Carbon on November 10, Hydrogen on November 12, and Neon on November 18. Google announced it through a blog post signed by Koray Kavukcuoglu, senior vice president at Google DeepMind and Google’s chief AI architect.

Google describes Neon as built for complex software engineering, knowledge work, and science workflows — the jobs where you need real reasoning depth but cannot wait for the slowest, most expensive model to finish. It is adapted from Carbon, which is why its scores land so close to the flagship’s, and it reads text, images, audio, video, PDFs, and spreadsheets inside a 1-million-token context window. One thing to know up front: it answers in text only. If your workflow needs image or video output, Neon is not the model for it.

There is no free API tier for Neon. It is paid tier only in the Gemini API, available through Google AI Studio and Android Studio. In the consumer apps, Pro and Ultra subscribers get it in the Gemini app (as the “Deeper Analysis” option) and in Google Antigravity, where it is also the default model for new conversations on plans that include it. It is available in Gemini for Business, it went generally available in Gemini Enterprise on launch day, and it shows up in Cursor as well. Its knowledge runs to March 2026.

One quiet change worth noting if you use the Gemini app: Neon has replaced Argon in the app for Pro and Ultra subscribers. Argon is still listed on the API pricing page at the same $2/$10 intro rate as Carbon, doubling to $4/$20 on December 1, but in the app it has stepped aside. Google’s naming scheme — elements in no particular order — has become a running joke among developers, and Neon, the tenth element, arriving right after Hydrogen, the first, does nothing to help.

The price table that matters: today vs December 1

Here is the whole Gemini 4 family, from Google’s Gemini API pricing page, as of November 19, 2026. These are standard-tier rates per million tokens, in US dollars. Output prices include thinking tokens — there is no separate thinking charge, which matters a lot for a reasoning model, because the tokens it spends thinking are billed at the full output rate.

Model Input (through Nov 30) Output (through Nov 30) Input (from Dec 1) Output (from Dec 1) Cached input Free API tier
Gemini 4 Hydrogen $1.50 $7.50 $1.50 $7.50 $0.075 Yes
Gemini 4 Neon $3.00 $15.00 $3.00 $15.00 $0.15 No
Gemini 4 Carbon $2.00 $10.00 $4.00 $20.00 $0.10 → $0.20 No
Gemini 4 Argon $2.00 $10.00 $4.00 $20.00 $0.10 → $0.20 No

Two things jump out. First, until November 30, Neon is 50% more expensive per token than Carbon. From December 1, it is 25% cheaper. Second, the cached-input discount is 95% across the board — Neon’s cached input is $0.15 per million against $3.00 uncached — so if your workload reuses long prompts or system instructions, caching changes the math more than the model choice does. Batch API requests are half the standard price everywhere, which puts Neon at $1.50/$7.50 in batch mode.

Now the worked math, because list prices do not pay the bill — token mixes do. Take a typical coding-agent task: 100,000 input tokens and 50,000 output tokens.

  • Neon: $0.30 input + $0.75 output = $1.05
  • Carbon, through Nov 30: $0.20 + $0.50 = $0.70
  • Carbon, from Dec 1: $0.40 + $1.00 = $1.40
  • Hydrogen: $0.15 + $0.375 = $0.53

Scale that to a month of serious use — 1 billion input tokens and 500 million output tokens:

  • Neon: $3,000 + $7,500 = $10,500
  • Carbon, through Nov 30: $2,000 + $5,000 = $7,000
  • Carbon, from Dec 1: $4,000 + $10,000 = $14,000

So a team running that workload on Carbon today and doing nothing will see its bill double on December 1, from $7,000 to $14,000. Moving the same workload to Neon holds it at $10,500 — a $3,500-a-month saving against the December Carbon price. And if half of that billion input tokens are cached prompts, Neon’s input cost drops from $3,000 to $1,575 (500 million at $3.00 plus 500 million at $0.15), bringing the monthly total to $9,075. Cache hit rate is the cheapest optimization in this whole table, and it is free to set up.

One honest caveat before you budget on these numbers: Neon writes about 15% more output tokens per task than Carbon, according to Artificial Analysis. Since output tokens are the expensive half of the bill, part of Neon’s price advantage is eaten by its own verbosity. The saving is real — it comes purely from the lower token prices — but it is smaller than the list-price gap suggests. Budget by cost per task, not by cost per token.

Is Neon actually better than Carbon?

This is where you have to separate two kinds of numbers: what Google says about its own model, and what independent testers measured. They do not agree, and the gap between them is the most useful thing in this story.

Google’s launch post says Neon beats Claude Opus 5.5 on several evaluations, including coding and agentic work, and that it tops the Artificial Analysis Coding Agent Index. It also quotes Glean’s AI product lead, Thai Tran, saying Neon “matched or exceeded Gemini 4 Carbon on overall quality at 3x the speed and 60% of the cost.” Those are Google’s claims, published in Google’s own announcement, and the SiliconANGLE coverage of the launch repeats the “three times as fast, 60% of the cost” framing. Treat them as the vendor’s pitch, not a measurement.

Here is what Artificial Analysis — the independent benchmark outfit — actually measured, published on launch day:

  • Intelligence Index: Neon scores 61 at its highest (xhigh) thinking level, one point ahead of Carbon at high (60), and third overall — behind Claude Fable 5.2 at 65 and GPT-6.2 Astra at 62, and ahead of Claude Opus 5.5 at 59.
  • Cost per task: $6.44 at xhigh, about 17% less than Carbon’s $7.76 (measured at Carbon’s standard $4/$20 pricing, not the intro rate).
  • Time per task: about 21 minutes against Carbon’s 36 — roughly 40% faster — because Neon generates around 110 output tokens per second against Carbon’s 57.
  • Coding Agent Index: Neon scores 58.6, ranked first, and in Google Antigravity it averages $6.86 per task, against $12.61 for Claude Opus 5.5 and $15.47 for Claude Fable 5.2.
  • CursorBench 4.1: Neon scores 65.4%, against 58.1% for Carbon and 57.3% for Hydrogen — a clear lead in live IDE editing.

Google also claims a 32% lower cost per task than Carbon at Neon’s default high thinking level. That figure and Artificial Analysis’s 17% figure are not contradictory — they are measured at different settings and different Carbon prices. Google’s 32% is at the default high level; the 17% is at xhigh, where Neon spends more tokens thinking. The honest summary is this: Neon matches or slightly beats Carbon on quality, finishes tasks meaningfully faster, and costs less per task — but the advantage is a solid margin, not the three-times-faster, 60%-cheaper story in the launch post. If you want the full comparison of the two models it trails, I broke down GPT-6.2 Astra vs Claude Fable 5.2 in a separate piece — Astra at 62 and Fable at 65 are the two models sitting above Neon on that index.

There is one more independent number worth knowing: on the Vals Index, Neon scores 70.9%. That figure comes from Google’s own materials, not an independent re-test, so I am listing it as a vendor-supplied result rather than a verified one.

The four thinking levels — and the one that changes your bill

Neon ships with four thinking levels: low, medium, high, and xhigh. High is the default. Xhigh is new to Gemini models with the recent Hydrogen release, and it is the setting that earned the 61 on the Intelligence Index. The thinking level is now a budget decision, not a quality toggle, because thinking tokens are billed as output at $15 per million.

That creates a real migration trap. Carbon defaults to medium thinking; Neon defaults to high. If you move a workload from Carbon or Argon to Neon and change nothing else, you are not just changing models — you are silently raising the thinking level, which raises both latency and output-token spend on routine prompts. Set the thinking level explicitly when you migrate. For everyday coding and document work, medium or high is the sensible default; reserve xhigh for the tasks where the extra reasoning is actually worth paying for, like a multi-file refactor or a research synthesis you will act on.

The practical rule: pick the lowest thinking level that still gets the job done, and test it on your own workload. The benchmark gap between high and xhigh is real but small — one index point — while the token bill difference is not.

Where you can use Neon today

Neon is live in more places than most launch-day models, because Google shipped it across its whole stack at once:

  • Gemini API (paid tier): call gemini-4-neon through Google AI Studio or Android Studio. No free tier — this is the one Gemini 4 model you cannot try without paying.
  • Gemini app: Pro and Ultra subscribers get Neon as the “Deeper Analysis” option.
  • Google Antigravity: available in the app, the CLI, and IDE extensions for paid Pro and Ultra subscribers, and it is the default model for new conversations on plans that include it. Note that the context window inside Antigravity is 512,000 tokens, not the full 1 million the API offers.
  • Gemini for Business and Gemini Enterprise: generally available from launch day.
  • Cursor and other editors: Neon is rolling out through the usual editor integrations.
  • Orchestrator (Neon): Antigravity’s new mode pairs the two models in one conversation — Neon plans and reviews, Hydrogen does the implementation work (code edits, package updates, test runs), and if Hydrogen gets stuck it hands control back to Neon to inspect the state, fix the plan, and continue. If you use Antigravity, this is the setup to try first: it is the cheapest way to get Neon’s judgment without paying Neon’s output rate for every keystroke.

For the wider context on the model Neon sits next to in the lineup, the site’s Gemini 4 Hydrogen launch coverage walks through the fast, cheap end of the family, including the free API tier Neon does not have.

Who should switch, who should stay

Here is the decision, stripped of the marketing:

Switch to Neon on December 1 if you run Carbon today. This is the clearest call in the whole launch. Your Carbon bill doubles on December 1 and Neon becomes 25% cheaper per token while matching Carbon on quality and finishing faster. There is no reason to keep paying flagship prices for flagship-adjacent work after the intro rate ends. Start the migration now, while Carbon is still cheap, so you can compare the two on your own workload before the price flip forces your hand.

Stay on Carbon through November 30 if your jobs are long-horizon and single-threaded. Carbon is still the model Google recommends for its hardest, longest tasks, and until November 30 it is also the cheaper one. If you have a big research run or a multi-day agent job already in flight, finish it on Carbon at the intro rate.

Use Hydrogen for the high-volume, low-stakes work. At $1.50/$7.50 with a free API tier, Hydrogen is half Neon’s price, and for straightforward tasks — drafting, summarising, simple code — the quality gap does not justify doubling the bill. The Antigravity Orchestrator setup formalises this: Neon thinks, Hydrogen types.

Do not switch just because of the leaderboard. Neon’s 61 versus Carbon’s 60 is a one-point gap on one index. If Carbon is already working for you and your workload is small, the migration is not worth an afternoon of your time. The price flip is the reason to move, not the benchmark.

Hold off if you need a free tier or image output. Neon has neither. Hydrogen has the free tier; if you need image generation, that is a different model family entirely.

The rough edges

No launch is clean, and Neon’s first two days had real friction worth knowing about before you build on it:

  • No free tier. Every other Gemini 4 launch in this cycle had some free path; Neon does not. You cannot try it without a paid API plan or a Pro/Ultra subscription.
  • Text-only output. It reads images, audio, video, and PDFs, but it only writes text. Plan accordingly.
  • A first-day bug in Antigravity. On November 18–19, some Ultra subscribers reported that Neon failed on xhigh with an “INVALID_ARGUMENT: Thinking level is unsupported” error, and that long chats were being cut to about 100,000 tokens. Google’s Antigravity team was responding to reports in its developer forum, but if you hit this, drop to high and wait for a fix rather than debugging your own setup. (Google AI Developers Forum thread)
  • March 2026 knowledge cutoff. Anything that happened after March 2026 is outside Neon’s training data. For a model launched in November, that is an eight-month gap — use search grounding for current facts.
  • The naming. Hydrogen, then Neon, skipping seven elements in between. It is cosmetic, but if you are writing model names into configs and documentation, double-check you typed the right element — “Carbon” and “Neon” are easy to confuse in a hurry, and they are about to cost very different amounts.

What this launch really says

Step back from the price table for a moment, because the launch tells you something about where Google is going. Four models in seven weeks, each one carving out a narrower slice of the workload spectrum, with pricing that only makes sense once you see the December rates. Google is not selling one model anymore — it is selling a ladder, and the ladder is priced so that the middle rung is the rational default for most serious work.

That is the real story of Neon. It is not the smartest model Google has ever shipped; Carbon still holds that title, and Fable 5.2 and Astra sit above it on the independent index. It is the model Google wants you to actually use every day: close enough to the flagship on quality, twice as fast, and — from December 1 — cheaper than the flagship it was adapted from. The launch-day price inversion, where the bigger model costs less than the mid-tier one, is a ten-day quirk. The December ladder is the product. Plan your December budget on it, not on November’s prices.

The pricing itself is the thing to verify before you act: every rate in this article comes from Google’s Gemini API pricing page, checked November 19, 2026, and the benchmark figures come from Artificial Analysis except where labelled as Google’s own claims. Prices change — Carbon’s doubling on December 1 is proof of that — so check the pricing page again before you commit a production workload.

Frequently Asked Questions

How much does Gemini 4 Neon cost?

Neon costs $3 per million input tokens and $15 per million output tokens on the standard paid tier, with cached input at $0.15 per million (a 95% discount). Batch API requests are half price at $1.50/$7.50. There is no free API tier, and the price is flat — unlike Carbon, it has no introductory rate and no scheduled change.

When does Gemini 4 Carbon get more expensive?

Carbon’s introductory price of $2 per million input tokens and $10 per million output tokens ends November 30, 2026. From December 1, Carbon costs $4 per million input tokens and $20 per million output tokens — double. Argon’s price doubles on the same date. Neon stays at $3/$15, which makes it 25% cheaper per token than Carbon from December 1.

Is Gemini 4 Neon better than Gemini 4 Carbon?

On independent benchmarks, Neon is roughly equal to Carbon with a speed advantage: it scores 61 on the Artificial Analysis Intelligence Index against Carbon’s 60, finishes tasks about 40% faster, and costs about 17% less per task at its highest thinking level. Google claims a bigger gap — 3x the speed at 60% of the cost — but that is the vendor’s own framing. Carbon remains Google’s recommended model for its hardest, longest-running tasks.

Where can I use Gemini 4 Neon?

Through the Gemini API (paid tier only) in Google AI Studio and Android Studio, in the Gemini app for Google AI Pro and Ultra subscribers, in Google Antigravity for paid Pro and Ultra subscribers, in Gemini for Business and Gemini Enterprise, and in Cursor. It is also the planning half of Antigravity’s Orchestrator mode, paired with Hydrogen for implementation work.

What are Neon’s four thinking levels?

Low, medium, high (the default), and xhigh. Higher levels mean deeper reasoning and more output tokens — which are billed at the full $15-per-million output rate. Xhigh earned the top benchmark score of 61, but for most everyday work, high or medium is the better value. Set the level explicitly when migrating from Carbon, which defaults to medium.

Should I switch from Carbon to Neon now or wait until December 1?

Start testing the migration now, while Carbon is still at its $2/$10 intro rate, and plan to switch production workloads by December 1, when Carbon doubles to $4/$20 and Neon becomes 25% cheaper per token. If your jobs are already running on Carbon and finishing before November 30, let them finish — Carbon is cheaper until the deadline.

Leave a Reply

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