What Is Gemini 4 Neon? Google’s Fast Frontier Model Explained
Gemini 4 Neon is Google’s answer to a practical question: what do you use when the everyday model is not quite enough and the flagship is more than you need? Announced on November 18, 2026, Neon is the new mid-tier member of the Gemini 4 family, built for complex coding, analytical reasoning and production-grade enterprise workflows. It costs $3 per million input tokens and $15 per million output tokens, runs about three times faster than Carbon on most tasks according to Google, and already leads the independent coding-agent ranking. This guide explains what Gemini 4 Neon is, where it fits in Google’s lineup, what it costs, how to use it, and where the launch claims deserve a second look. Benchmark figures are labelled throughout: Google’s own numbers are Google’s, and independent figures come from Artificial Analysis, which tests every model through the same harness.
What is Gemini 4 Neon?
Neon is a large language model from Google DeepMind, and the third model in the Gemini 4 series to be aimed at developers and businesses. Argon arrived on October 13 as the deep-reasoning flagship, Carbon followed on October 28 as the strongest model in the range for the hardest, longest engineering jobs, and Hydrogen completed the family from the other direction on November 12 as the fastest and cheapest option for everyday work. Neon occupies the space the first three left open: serious, production work that needs real reasoning depth, at a price and speed that let you run it all day.
Google’s own description of the three-way split is simple. Carbon is for the hardest long-horizon projects. Hydrogen is for fast everyday execution. Neon is for complex coding, analytical reasoning and production-grade enterprise workflows. The distinction is about the job, not a single leaderboard — and it is the key to choosing between them.
Under the hood, Neon shares the family architecture: a 1 million-token context window, enough to hold large repositories and long documents in full before work starts, and four configurable thinking levels — low, medium, high (the default) and xhigh — so developers can choose how much reasoning a task gets and pay only for the depth it needs. A quick code completion can run on low; a tangled debugging job or a multi-step analysis can go to xhigh.
Where Neon sits in the Gemini 4 family
| Model | Role | Price ($/1M input / output) | Artificial Analysis index | Best for |
|---|---|---|---|---|
| Gemini 4 Carbon | Strongest, long-horizon | $4 / $20 (after $2 / $10 intro) | 60 | Hardest engineering, cybersecurity |
| Gemini 4 Argon | Deep reasoning | $2 / $10 | 50 | Everyday engineering where depth matters |
| Gemini 4 Neon | Mid-tier frontier | $3 / $15, $0.15 cached input | 61 (xhigh) | Complex coding, enterprise workflows |
| Gemini 4 Hydrogen | Fast everyday execution | $1.50 / $7.50, free tier | 51 (high), 54 (xhigh) | Everyday coding, agentic execution, Search |
Two things stand out in that table. First, Neon is the highest-scoring Google model on the independent index — 61 at xhigh, one point ahead of Carbon’s 60 and seven ahead of Hydrogen’s 54 at the same setting. That is a striking result for a mid-tier model, though it needs the usual caution: the index measures a fixed set of evaluations, and a one-point lead over Carbon does not mean Neon is the stronger model on the hardest work. Second, the price ladder is now evenly spaced: $1.50 for Hydrogen, $3 for Neon, $4 for Carbon on input tokens. Google is selling a choice of speed against depth, with a model at each step.

Against rivals outside Google, the gap is smaller than it has been all year. Claude Fable 5.2 still leads the Artificial Analysis index at 65, and GPT-6.2 Astra sits second at 62 — but Neon at 61 is within a point of Astra while costing less than half as much per token: $3 against Astra’s $8 input price. For the full comparison of those two rivals, see our guide to GPT-6.2 Astra vs Claude Fable 5.2.
The benchmarks: independent ranking vs Google’s claims
Neon’s results split cleanly into two groups — what an independent lab measured, and what Google reported about its own model. Read them with that split in mind.
| Benchmark | Gemini 4 Neon | Source |
|---|---|---|
| Artificial Analysis Intelligence Index | 61, rank #3 (xhigh) | Independent (Artificial Analysis) |
| Vals Index | 70.9%, rank #2 | Independent |
| Agent Arena Index | 1,336, rank #2 | Independent |
| AA Coding Agent Index | 58.6, rank #1 | Independent (Artificial Analysis) |
| Terminal-Bench 4.1 | 65.2% | Google’s claim (default high) |
| FrontierCode 2.0 | 46.2% | Google’s claim (default high) |
| Terminal-Bench Science | 70.5%, best in class | Google’s claim |
| RiemannBench | 84.0% | Google’s claim |
| HLE-Diamond | 71.3% | Google’s claim |
| AutoBaxBench (hard tier, cyber) | 76%, state of the art | Google’s claim |
| WebExploitBench v1.2 (cyber) | 51.26%, state of the art | Google’s claim |
| SEC-bench Pro (cyber) | 90.5%, state of the art | Google’s claim |
| CWE-bench v2 (cyber) | 70.7%, state of the art | Google’s claim |
The independent results are the ones that matter most, and they tell a consistent story. Neon is a top-three model overall, second on two of the live evaluation indexes, and first on the coding-agent index — the first time a Google model has topped an Artificial Analysis coding ranking. The coding result is the headline: on the task that matters most to developers — whether an agent can actually complete coding work in a real repository — Neon now leads the field.
Google’s own figures are stronger still, and they are where the caution belongs. Terminal-Bench 4.1 at 65.2% and FrontierCode 2.0 at 46.2% are both measured at the default high thinking level, and the science results — 70.5% on Terminal-Bench Science, 84.0% on RiemannBench, 71.3% on HLE-Diamond — are described by Google as best in class. No outside lab had independently reproduced those numbers at launch, and vendor benchmarks are chosen by the vendor. The cyber scores are the most striking of all, and also the most restricted in who can verify them: the model is rolling out to trusted defenders through Google’s Fairwind Program, not to the general public.
Two cost claims also come from Google alone. Neon is, the company says, about three times faster than Carbon on most tasks, and 32% cheaper per task than Carbon at default high thinking — roughly 50–80% cheaper per task than the top competing frontier models. The speed figure comes from SiliconANGLE’s coverage of the launch as well as Google’s own post, but the cost-per-task comparison has not yet been independently reproduced. Treat it as Google’s claim until independent figures appear.
Gemini 4 Neon price and how to access it
Neon costs $3 per million input tokens, $0.15 for cached input and $15 per million output tokens through the Gemini API, flat across its entire 1 million-token context window. The cached-input price is the number that changes the economics: at a 95% discount on the input rate, long agent sessions that re-read the same repository or document set cost far less than the headline $3 suggests. For comparison, Gemini 4 Hydrogen costs $1.50 input and $7.50 output, and Gemini 4 Carbon costs $4 input and $20 output at its standard rate. Neon sits exactly between them on input cost and, Google says, delivers 32% lower cost per finished task than Carbon.
Unlike Hydrogen, Neon has no free tier in the Gemini API at launch. Access from November 18 runs through four channels:
- Developers: the Gemini API in Google AI Studio, on the paid tier.
- Businesses: Gemini for business, through the API or the model picker.
- Google Antigravity: the agent-first development platform, for paid Pro and Ultra subscribers.
- Consumers: the Gemini app, through the model dropdown, for Pro and Ultra subscribers. Tasks inside Gemini can now run on Neon as well.
Google has not published a country-by-country rollout schedule beyond the global launch, and availability in specific products can still vary by region. It has also not said when Neon will reach Vertex AI, the cloud platform many enterprises buy through.
What Neon can do
- Complex coding. Neon’s home ground is production code: resolving issues in large repositories, debugging, rapid iteration and answering questions about a codebase, with the top-ranked coding-agent score to back it. On FrontierCode 2.0, which tests whether code changes meet the standards for merging into real projects, Google reports 46.2% at default high thinking.
- Analytical reasoning. The model is built for multi-step analysis — working through data, weighing evidence and producing a finished deliverable rather than a single-turn answer. Its second-place ranking on the Vals Index, which tests professional agent work, and on the Agent Arena Index both point in this direction.
- Enterprise workflows. Google pitches Neon as the model for production-grade business use: turning dense documents, spreadsheets and reports into structured outputs, running the multi-step tasks that make up a working day, and doing it fast enough and cheaply enough to leave running. For background on how those agent workflows operate, see our guide to agentic AI.
- Science. Google reports best-in-class results on three science benchmarks at default high thinking: 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond. The results suggest a model that can interpret real data and run experiments to infer how a system works, though all three figures are Google’s own.
- Cybersecurity — for defenders only. Neon is, Google says, state of the art on four cyber benchmarks: 76% on the hard tier of AutoBaxBench, 51.26% on WebExploitBench v1.2, 90.5% on SEC-bench Pro and 70.7% on CWE-bench v2. Access is restricted: the model is rolling out to trusted defenders through the Fairwind Program in the coming days, and it is aimed at verified security work rather than general use.
- The Gemini app. For Pro and Ultra subscribers, Neon is now a choice in the app’s model dropdown, and tasks inside Gemini can run on it — the first time a mid-tier frontier model has been offered directly to consumers in the app rather than only through the API.

Safety: low fabrication, zero cheating, zero injection success
Google’s safety claims for Neon are unusually specific. The company reports that Neon has the lowest fabrication rate yet recorded in internal testing in Google Antigravity — fabrication being the model’s tendency to invent unsupported claims — a 0.0% cheating rate on Terminal-Bench 4.0, and a 0.0% attack success rate on the Gray Swan indirect-prompt-injection benchmark at the highest thinking level. Indirect prompt injection is the attack where a model is tricked by instructions hidden in data it reads, such as a web page or a document; a zero success rate at the highest reasoning setting, if it holds outside Google’s lab, is a meaningful result for anyone deploying agents that browse the web or read files.
All three figures are internal results, and the safety case will be tested by outside researchers in the weeks after launch. The cyber capabilities also cut both ways: a model this strong at finding and exploiting vulnerabilities is a powerful tool for defenders, and a powerful tool in the wrong hands. That is why Google is gating cyber access through Fairwind rather than opening it to everyone.
Limits and open questions
- Vendor benchmarks. The biggest wins — Terminal-Bench 4.1, FrontierCode 2.0, the science tests and all four cyber benchmarks — are Google’s own tests, run on its own configurations. The independent index still ranks Claude Fable 5.2 first and GPT-6.2 Astra second. Treat the 3x-faster and 32%-cheaper headlines as Google’s claims until independent cost-per-task figures appear.
- No free tier. Unlike Hydrogen, Neon is paid-only in the API at launch. Developers who want to try the family without paying will start with Hydrogen, not Neon.
- Not in Search, not for free users. Hydrogen reached every Google Search user and every free Gemini app user; Neon is limited to paid tiers and Pro/Ultra subscribers. Most people who use Google’s products will never meet this model directly.
- The one-point lead. Neon’s 61 on the independent index is one point ahead of Carbon’s 60 and one behind Astra’s 62. Index points are not a ladder of overall intelligence — a narrow lead on one set of evaluations does not mean Neon is the stronger model on the hardest work, where Carbon’s long-horizon results still stand.
- Vertex AI and regional availability. Google has said nothing about when Neon reaches Vertex AI, and nothing about a country-by-country schedule. Enterprise buyers who purchase through Google Cloud will need to wait for that announcement.
- The price war cuts both ways. Neon is cheaper per task than Carbon, and Microsoft launched seven cheaper models the same week. If the frontier labs keep cutting prices to win developers, the cheapest model may soon be good enough for most work — which is good for users and bad for anyone betting on premium pricing.
Who should use Gemini 4 Neon?
Consider Neon if you build or run production software and measure cost per finished task rather than cost per token. At $3 per million input tokens with 95%-discounted caching, it is the best value among the frontier models for complex coding and enterprise agent work, and its independent coding-agent ranking is the strongest evidence Google has produced that the value is real.
Consider Hydrogen first if your work is everyday coding, routine debugging or agentic execution where speed and cost matter most. It costs half as much per token, has a free tier, and reaches every Search user — and at 54 on the independent index at xhigh it is closer to Neon than the price gap suggests. Our guide to Gemini 4 Hydrogen explains the trade-off.
Look elsewhere if you need the single highest independent score: Claude Fable 5.2 still leads the Artificial Analysis index at 65, and our guide to Claude Fable 5.2 explains what it does differently. Use Carbon instead if the work is a multi-day engineering project or security-sensitive — Neon is faster and cheaper, but Carbon remains Google’s strongest model for the hardest jobs. Skip Neon too if you only use the free tier: it is not there yet.
Frequently asked questions
What is Gemini 4 Neon?
Gemini 4 Neon is Google’s new mid-tier frontier model, launched November 18, 2026, built for complex coding, analytical reasoning and production-grade enterprise workflows. It costs $3 per million input tokens and $15 per million output tokens, and scores 61 on the Artificial Analysis Intelligence Index at xhigh thinking.
How much does Gemini 4 Neon cost?
$3 per million input tokens, $0.15 for cached input and $15 per million output tokens through the Gemini API. There is no free tier at launch. Google says it costs 32% less per task than Carbon and 50–80% less per task than the top competing frontier models.
Is Gemini 4 Neon better than Carbon?
On the independent Artificial Analysis index, Neon scores 61 at xhigh against Carbon’s 60 — a one-point lead. But Carbon remains Google’s strongest model for the hardest long-horizon projects, and Neon is faster and cheaper rather than more capable at the top end. The choice is about the job, not the score.
How is Neon different from Hydrogen?
Hydrogen is the fast, cheap everyday model at $1.50 and $7.50 with a free tier; Neon is the mid-tier frontier model at $3 and $15 with no free tier, built for harder coding and enterprise work. Neon scores higher on the independent index — 61 at xhigh against Hydrogen’s 54 — but costs twice as much per token.
How do I use Gemini 4 Neon?
Through the Gemini API in Google AI Studio on the paid tier, through Gemini for business, in Google Antigravity for paid Pro and Ultra subscribers, or from the model dropdown in the Gemini app for Pro and Ultra subscribers. Tasks inside Gemini can also run on Neon.
Is Gemini 4 Neon safe?
Google reports the lowest fabrication rate yet in its internal Antigravity testing, a 0.0% cheating rate on Terminal-Bench 4.0, and a 0.0% attack success rate on the Gray Swan indirect-prompt-injection benchmark at the highest thinking level. All three are internal results that outside researchers have not yet verified.