November 21, 2026

Gemini 4 Neon Explained: Google’s Fast New AI Model — and the December 1 Price Flip That Makes It Cheaper

Glowing neon-blue orb of light above a computer chip

Google has now released four Gemini 4 models in about seven weeks, and the newest one — Gemini 4 Neon, launched November 18, 2026 — is the most confusing of the lot. It is nearly as smart as the flagship Carbon, nearly twice as fast in early tests, and yet for the next twelve days it costs more per token than Carbon does.

Then, on December 1, the comparison flips. Carbon’s introductory price doubles, and Neon becomes the cheaper model by 25%.

This guide explains where Neon sits in the Gemini 4 family, what it actually costs per finished job, how its benchmarks stack up against Carbon, Claude Fable 5.2 and Claude Opus 5.5, and who should switch to it now — and who should wait until December.

Gemini 4 Neon at a glance

LaunchedNovember 18, 2026 (Google DeepMind blog, Koray Kavukcuoglu)
Price$3 per million input tokens, $15 per million output tokens; cached input $0.15 per million (95% discount)
Context window1 million tokens (text, images, audio and video)
Thinking levelsFour — low, medium, high and xhigh
Artificial Analysis Intelligence Index61 at xhigh — #3 overall; 60 at the default high setting
Vals Index#2, 70.9%
Coding Agent Index#1, 58.6
Where to get itGemini API and Google AI Studio, Gemini for business, Google Antigravity (Pro and Ultra), Gemini app (Pro and Ultra)
Free tierNone — the Gemini app requires a Google AI Pro ($19.99 a month) or Ultra subscription
The catchThrough November 30, Carbon’s introductory $2/$10 rate makes Carbon cheaper per token; from December 1, Neon is 25% cheaper

Where Neon sits in the Gemini 4 family

Google now has four Gemini 4 models, and the names tell you almost nothing about which one to pick. Here is the lineup as it stands on November 21:

ModelRolePrice per 1M tokens (input / output)Artificial Analysis Index
Gemini 4 HydrogenFast everyday work$1.50 / $7.5051 high / 54 xhigh
Gemini 4 NeonThe middle — complex work, done fast$3 / $1560 high / 61 xhigh
Gemini 4 CarbonHardest long-horizon projects$2 / $10 until Nov 30, then $4 / $2060
Gemini 4 ArgonBeing replaced by Neon in the Gemini app$2 / $10 until Nov 30, then $4 / $2050

Neon sits between Hydrogen, the small fast workhorse we covered in our Gemini 4 Hydrogen explainer, and Carbon, the flagship we covered when it opened to paid users on November 10. Google’s pitch is that Neon brings “frontier-level intelligence with fast end-to-end task completion” — Carbon-level quality at close to Hydrogen speed. Early users back the speed claim: Glean’s AI product lead told Google that Neon “matched or exceeded Gemini 4 Carbon on overall quality at 3x the speed and 60% of the cost” in their early evaluations.

Neon also replaces Argon in the Gemini app, so if you have been using Argon, Neon is now the model in that slot.

What Neon costs per finished job

Token prices are only half the story. What matters is what one finished job costs, and that depends on how many tokens the job uses. Take a typical coding job — say, fixing a bug in an unfamiliar codebase — that uses 100,000 input tokens and produces 20,000 output tokens.

Example 1 — one job, four models. Neon: 100,000 input tokens at $3 per million is $0.30, plus 20,000 output tokens at $15 per million is $0.30 — $0.60 total. Carbon from December 1 ($4/$20): $0.40 plus $0.40 — $0.80. Claude Fable 5.2 ($8/$40): $0.80 plus $0.80 — $1.60. Claude Opus 5.5 ($4/$20): $0.80. Hydrogen ($1.50/$7.50): $0.30. Neon costs a third less than Fable 5.2 and Opus 5.5 on this job, and a quarter less than Carbon’s December price.

Example 2 — a month of jobs. Run 100 of those jobs a month and the bill is $60 on Neon, $80 on Carbon at its December price, $160 on Fable 5.2 and $30 on Hydrogen. Neon saves $100 a month against Fable 5.2 — $1,200 a year — for a model that scores higher than Fable on the independent Artificial Analysis index. Against Carbon, the saving is $20 a month, or $240 a year.

Example 3 — the independent cost-per-task check. Artificial Analysis measures what a standard set of tasks actually costs, and it puts Neon at $6.44 per task at its highest thinking level, against $7.76 for Carbon at its undiscounted price — $1.32 less per task, or about 17% cheaper. Over 100 tasks, that is $132 saved. Inside Google’s Antigravity coding tool, an average coding task costs $6.86 with Neon, against $12.61 for Opus 5.5 and $15.47 for Fable 5.2.

Example 4 — the caching discount. Most of a long coding session is the model re-reading code it has already seen. Neon charges $0.15 per million cached input tokens — a 95% discount on the $3 fresh-input rate. A job with 1 million cached tokens, 100,000 fresh input tokens and 20,000 output tokens costs $0.15 plus $0.30 plus $0.30 — $0.75. If your workflow reuses a lot of context, caching is where the real savings live.

The December 1 price flip — and why it matters now

Here is the quirk that makes the timing matter. Carbon is still on its introductory price of $2 per million input tokens and $10 per million output tokens until November 30. Neon launched at a flat $3 and $15. So for the next twelve days, Neon costs 50% more per token than Carbon.

On December 1, Carbon’s price doubles to $4 and $20, and Neon — whose price does not change — goes from costing 1.5 times as much as Carbon to costing 0.75 times as much.

Through Nov 30From Dec 1
Carbon, 100K in / 20K out job$0.40$0.80
Neon, same job$0.60$0.60
Neon vs Carbon50% more expensive25% cheaper

If you do not want to wait until December, the Batch and Flex tiers — where you accept slower processing for a lower bill — cut Neon to $1.50 and $7.50 per million tokens, the same as Hydrogen’s standard rate.

The practical advice is simple: if you are already running Carbon workloads, keep them where they are until November 30, then move most of them to Neon on December 1 and keep Carbon only for the jobs where Neon visibly struggles.

How Neon benchmarks

Google’s own benchmark numbers are vendor-reported, so the independent scores matter more. Here is what the launch announcement and the Artificial Analysis leaderboard show:

BenchmarkNeonCarbonFable 5.2 MaxOpus 5.5
Artificial Analysis Intelligence Index61 (xhigh), 60 (high) — #3 overall606559
Vals Index (professional tasks)70.9% — #2———
Artificial Analysis Coding Agent Index58.6 — #1———
Terminal-Bench 4.165.2%———
FrontierCode 2.046.2% (high)—49.1%45.5%
Terminal-Bench Science70.5% — best in class———
HLE-Diamond71.3% — best in class———
RiemannBench84.0% — best in class———

Three things stand out. First, Neon edges past Carbon on the general intelligence index — 61 against 60 — while running at roughly two to three times the speed in early tests. Second, it ranks #1 on the Coding Agent Index, ahead of both Claude models, and #2 on the Vals Index for professional tasks like finance, accounting and spreadsheet modelling, where it ranks #1 on most Surge AI benchmarks. Third, it is the only model in this set that is best in class on Terminal-Bench Science, HLE-Diamond and RiemannBench at the same time.

Neon also ranks #2 on steerability — how cleanly it adapts when you correct it mid-task — on the new Agent Arena Index, and Google reports its lowest fabrication rate yet in internal Antigravity testing. On cybersecurity benchmarks it posts 51.26% on WebExploitBench, 90.5% on SEC-bench Pro and 70.7% on CWE-bench v2.

The honest caveat: one scoreboard is never the whole story. Claude Fable 5.2 still leads the general index at 65, and our cost-per-finished-task comparison showed that the cheapest model per token is often not the cheapest per finished job — it depends on how many tries the model needs.

Who should use Neon — and who should not

Use Neon if: you run coding agents or long document work all day and your bill is mostly output tokens; you are on Google AI Pro or Ultra and want the strongest model in the Gemini app; or you are choosing between Carbon and a Claude model and want the cheaper per-task option without giving up quality.

Wait until December 1 if: you are already on Carbon’s introductory rate and your workloads are running fine. There is no reason to pay 50% more per token for the next twelve days.

Do not switch if: you mostly do short, simple jobs — Hydrogen at $1.50/$7.50 does those for half the price of Neon. Or if you need the very top score on the general index, where Fable 5.2 still leads — though you will pay $1.60 a job against Neon’s $0.60 for that edge.

Watch out for: the missing free tier. Neon has no free tier on the Gemini API, and in the Gemini app it is only for paying Pro and Ultra subscribers. If you are on the free tier, nothing changes — you will not see Neon in your model menu.

What to watch next

The next test for Neon is not a benchmark but a month of real usage: whether developers actually move their Carbon workloads over on December 1, and whether the 3x-speed claim holds up outside Google’s own tests. The other date to watch is December 2, when the EU AI Act’s transparency deadline arrives — a rule that affects how all of these models label their output. For now, Neon is the fourth Gemini 4 model in seven weeks, and the pace alone says something: Google is shipping a new flagship-class model roughly every fortnight, and each one is cheaper per task than the last.

Checked against primary sources on November 21, 2026.

Frequently asked questions

What is Gemini 4 Neon?

Gemini 4 Neon is Google’s fourth Gemini 4 model, launched November 18, 2026. It sits between Hydrogen, the fast everyday model, and Carbon, the flagship for the hardest projects, offering near-Carbon quality at roughly two to three times the speed in early tests.

How much does Gemini 4 Neon cost?

Neon costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million — a 95% discount. Batch and Flex tiers cut it to $1.50 and $7.50. There is no free tier on the API.

Is Gemini 4 Neon cheaper than Carbon?

From December 1, yes — Neon is 25% cheaper per token than Carbon, whose price doubles from $2/$10 to $4/$20. Through November 30, Carbon’s introductory rate makes it the cheaper option, so most Carbon users should wait until December to switch.

How does Neon rank on benchmarks?

Neon scores 61 on the Artificial Analysis Intelligence Index at its highest thinking level, ranking third overall behind larger, slower models. It ranks second on the Vals Index, first on the Coding Agent Index, and best in class on Terminal-Bench Science, HLE-Diamond and RiemannBench.

Where can I use Gemini 4 Neon?

Neon is available in the Gemini API and Google AI Studio, Gemini for business, Google Antigravity for Pro and Ultra subscribers, and in the Gemini app for Google AI Pro ($19.99 a month) and Ultra subscribers. It replaces Argon in the Gemini app.

Should I switch from Claude Fable 5.2 or Opus 5.5 to Neon?

It depends on your workload. Neon costs about a third less per finished job than either Claude model on a typical coding task, and it outscores both on the Coding Agent Index. But Fable 5.2 still leads the general intelligence index, so keep it for your hardest reasoning work.

Sources