⚡ Quick verdict

We may earn a commission if you buy through our links, at no extra cost to you.

Google launched Gemini 4 Neon on November 18, 2026 — the fourth Gemini 4 model in about seven weeks, after Argon, Carbon and Hydrogen. The pitch is a balancing act: frontier-level intelligence at Flash-like speed, for the complex work that shows up every day rather than once a quarter. If Hydrogen is the fast daily driver and Carbon is the heavyweight for the hardest problems, Neon is the model Google wants you to reach for in between — the one that runs your coding agents all day without the flagship bill.

The numbers Google published are genuinely strong. Neon scores 61 on the Artificial Analysis Intelligence Index — third in the world, behind only Claude Fable 5.2 (65) and GPT-6.2 Astra (62) — and it tops the Coding Agent Index at 58.6. It does this at $3 per million input tokens and $15 per million output tokens, with roughly twice the output speed of Carbon, according to Artificial Analysis. On paper, that is the best price-to-intelligence ratio Google has ever shipped.

But launch-day numbers deserve a sceptical read, and this review gives you one. Some of Neon's headline figures come from Google's own test setups — most notably the SWE-sweep score, which triples when the model runs inside Google's own Antigravity environment. Other figures, like the 65.2% on Terminal-Bench 4.1, come from public methodologies anyone can rerun. This review is based on Google's launch post and model card, the Artificial Analysis index snapshots, and first-day independent coverage from November 18–19, 2026. We have not run our own benchmarks — nobody outside Google has had the model for more than a day — so where a claim comes only from Google, we say so, and where something is still unconfirmed, we flag it. Pricing and availability change fast; check the official site before you build on any figure here.

What is Gemini 4 Neon?

Gemini 4 Neon is a general-purpose AI model adapted from Gemini 4 Carbon, Google's flagship. “Adapted from” is Google's own phrasing: Neon inherits much of Carbon's capability but is tuned for speed and for the high-frequency work that dominates real developer days — complex coding, analytical reasoning, enterprise knowledge work and science. It is not a smaller model in the usual sense; it is a Carbon-derived model rebalanced for fast, repeated use.

It accepts text, images, audio, video and PDFs as input and returns text, with a knowledge cutoff of March 2026 — so anything newer needs live search or your own data. The model ID developers will use is gemini-4-neon, and it runs on the paid tier of the Gemini API only: there is no free tier, which is a real difference from Hydrogen and worth knowing before you plan a weekend experiment.

In the Gemini app, Neon replaces Argon and takes the “Deeper analysis” slot in the model picker, between Hydrogen (“Everyday help”) and Carbon (“Code & complex work”). Google's 3.1 Pro model is going away as part of the same reshuffle. For subscribers, that means the choice is now explicit: everyday questions go to Hydrogen, deeper analysis to Neon, and the hardest code and complex work to Carbon.

Gemini 4 Neon pricing: genuinely cheap, with one catch

Neon costs $3 per million input tokens and $15 per million output tokens on the standard paid tier of the Gemini API. Cached input — prompt text Google has already stored — costs $0.15 per million tokens, a 95% discount on the input rate, and only implicit caching is supported; explicit caching is not available for this model. Batch and Flex API tiers are listed at half the standard rate, according to launch coverage. Prices are from launch week, November 2026 — check the official site for current pricing.

The honest way to read the price is against its siblings. Hydrogen costs $1.50 in and $7.50 out — exactly half of Neon. Carbon costs $2 in and $10 out on its introductory promotion through November 30, 2026, then rises to $4 and $20 from December 1. That creates a strange twelve-day window: until November 30, Neon is 50% more expensive per token than the flagship Carbon. From December 1, Neon becomes 25% cheaper than Carbon's standard rate. If you are evaluating the family this week, compare against December prices, not today's — the promotional window distorts every comparison.

In absolute terms, the prices are easy to feel. A request with 50,000 uncached input tokens and 6,000 billed output tokens — including thinking tokens — costs about $0.24 at standard rates. A coding agent that burns a million input tokens and 200,000 output tokens in a day costs $3 plus $3, or $6 a day. A small team running agents all month might use 50 million input tokens and 10 million output tokens: $150 plus $150, or $300 a month. Output is where the money goes — at five times the input rate, long generations dominate the bill, so prompts that ask for concise answers save real money.

ModelPositionInput / 1MOutput / 1MIntelligence Index
Gemini 4 NeonMid-tier, high-frequency work$3.00$15.0061 (xhigh)
Gemini 4 HydrogenFast everyday work$1.50$7.5051 (high) / 54 (xhigh)
Gemini 4 CarbonFlagship, hardest problems$2.00 promo / $4.00 from Dec 1$10.00 promo / $20.00 from Dec 160
Gemini 4 ArgonDeep reasoning sibling$2.00 promo / $4.00 from Dec 1$10.00 promo / $20.00 from Dec 150
GPT-6.2 SolOpenAI everyday model$2.00$10.0057

Carbon and Argon promotional pricing runs through November 30, 2026; standard rates apply from December 1. Prices as published November 18–19, 2026 — check the official site for current pricing.

Thinking levels: low, medium, high, xhigh

Like Hydrogen, Neon ships with four thinking levels: low, medium, high (the default) and xhigh. The thinking level controls how much internal reasoning the model does before answering. Low is for fast, high-volume work where the answer is mostly retrieval or formatting. Medium suits interactive coding. High is the default and the right setting for most work. Xhigh is reserved for hard, multi-step problems where high falls short — it uses more tokens and takes longer, so it costs more per task.

The thinking level is not a cosmetic setting; it moves the benchmark scores. On the Artificial Analysis Intelligence Index, Neon scores 60 at its default high level and 61 at xhigh — one point buys the #3 world ranking. On the Coding Agent Index, the xhigh setting is what takes Neon to the top of the table. That means the model you actually get depends on the setting you choose: at default settings Neon is a 60; pushed to xhigh it is a 61. For most users the default is the honest number to remember, and xhigh is the setting to reach for when a task earns it.

There is a practical consequence for API users: thinking tokens are billed as output, and they share the same 1,048,576-token context window as input and output. A long reasoning chain eats into the same budget as your prompt and your answer. Google suggests using low for the trivial steps in an agent loop, medium for interactive work, and xhigh only where accuracy is worth the wait. Treat the thinking level as a first-class setting in your configuration — the difference between low and xhigh on a hard task can be the difference between a fast wrong answer and a slow right one.

Speed and benchmarks: fast, and near the top where it counts

Google's headline speed claim is that Neon delivers Carbon-level quality at Flash speeds. Artificial Analysis measured roughly 110 output tokens per second for Neon against about 57 for Carbon — about twice as fast. For agentic work, where a model may take dozens of steps in sequence, that speed compounds: a coding agent that takes 10 minutes on Carbon finishes in about 5 on Neon, and the shorter the wait, the more loops a developer is willing to run.

The benchmark Google leads with is the Artificial Analysis Coding Agent Index, where Neon scores 58.6 — first overall. On Terminal-Bench 4.1, a test of agentic coding in real terminal environments, Neon scores 65.2%, against 65.8% for Claude Opus 5.5 and 61.5% for GPT-6.2 Sol. Opus edges Neon by a fraction of a point on that one; Neon leads the composite. On FrontierCode 2.0, Neon scores 46.2% at high — again slightly behind Opus. The pattern is consistent: Neon is at or near the top of every coding benchmark, first on the composite, a fraction behind Opus on two individual tests.

Beyond coding, the numbers stay strong. Neon scores 95.2% on GraphWalks, a long-context test, against 80.1% for GPT-6.2 Sol and 66.8% for Claude Opus 5.5 — a wide lead that suggests the million-token window is genuinely usable, not just a specification. On professional knowledge work it scores 70.9% on the Vals Index (second overall) and 75.8% on Agents' Last Exam, with 56.3% on AutomationBench. In science and maths it records 70.5% on Terminal-Bench-Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond. Google's own comparison table shows Neon ahead of Opus 5.5 and GPT-6.2 Sol on most of the 18 benchmarks it publishes — but remember that table is Google's own selection, and one independent analysis found Neon winning 52% of rows, tying 33% and losing 14% once third-party boards are included. The honest summary: Neon is a top-three model by any measure, and the best coding-agent model on the composite index — but it does not beat Opus on every test.

One more Google claim is worth a careful read: at its default high thinking level, Neon costs 32% less per completed task than Carbon, according to Google. Artificial Analysis measured a similar shape independently: about $6.44 per task for Neon at xhigh against $7.76 for Carbon at high on standard pricing — roughly 17% less. The two figures measure different settings, so they are not directly comparable, but they point the same way: Neon does comparable work for meaningfully less money than the flagship. That is the core of its case.

BenchmarkGemini 4 NeonClaude Opus 5.5GPT-6.2 Sol
Artificial Analysis Intelligence Index61 (xhigh)5957
Terminal-Bench 4.165.2%65.8%61.5%
SWE-sweep, provider harness39.0% (Antigravity)10.1% (Claude Code)8.3% (Codex)
SWE-sweep, shared mini-SWE harness12.8%3.6%11.6%
GraphWalks (long context)95.2%66.8%80.1%

Scores as published in Google's launch materials and Artificial Analysis snapshots, November 18–19, 2026. SWE-sweep figures depend heavily on the harness used — see the next section.

The SWE-sweep number needs a warning label

The single most quoted Neon number — 39.0% on SWE-sweep — needs a warning label, because the same model scores 12.8% on the same benchmark in a different harness. A harness is the software loop around a model: the tools it gets, the feedback it receives, the way multi-step work is managed. Google's 39.0% was measured inside Google Antigravity, which can send subagents across a repository in parallel. In the shared mini-SWE harness that every model can use, Neon scores 12.8%. Opus shows the same pattern in reverse: 3.6% in the shared harness, 10.1% inside Claude Code. Sol actually falls: 11.6% shared, 8.3% inside Codex.

What this means is that the headline comparison — Neon 39.0% against Opus 10.1% and Sol 8.3% — compares agent products as much as it compares models. On the shared harness, where the playing field is level, Neon still leads (12.8% vs 11.6% for Sol and 3.6% for Opus), but the gap shrinks from 29 points to about 1 point over Sol. Neither number is wrong; they measure different things. If you use Antigravity, the 39.0% is the relevant figure. If you use anything else, the 12.8% is. Any article that quotes one without the other is telling half the story.

This is also why the Coding Agent Index composite matters more than any single row: it aggregates many tasks and harnesses, and there Neon leads outright at 58.6. Single benchmarks are marketing; composites and your own tests are evidence.

Antigravity: Neon plans, Hydrogen executes

Neon's most distinctive feature is not a benchmark but a job description: inside Google Antigravity, Neon can act as an Orchestrator — planning a long-running job, splitting it into smaller tasks, and handing those tasks to cheaper Hydrogen subagents working in parallel. Carbon previously handled that planning role; Neon takes it over at a lower reported cost per completed task. The idea is simple economics: one expensive planner directing many cheap executors is cheaper than one expensive model doing everything itself.

In practice, this is the agentic-coding pattern that matters most for real teams: a big refactor, a migration across services, a bug that touches five files in three packages. The planner holds the whole picture; the executors do the local work. Google's claim is that this pattern is what drives the Antigravity SWE-sweep number — the subagent tool lets agents search different parts of a codebase in parallel. Whether that justifies the headline figure is a fair debate; what is not debatable is that the planner-executor pattern is now a first-class feature of Google's coding environment, and Neon is the model built for it.

Access is the catch. Antigravity requires a paid Google AI Pro or Ultra subscription, and enterprise access goes through Google's business plans. If you already pay for Pro or Ultra, Neon in Antigravity is included in what you already pay — the value case is immediate. If you do not, the subscription is the price of admission, and the API is the cheaper way to try the model first.

Where you can use Gemini 4 Neon

At launch, Neon is available in the Gemini app for AI Pro and AI Ultra subscribers, through the Gemini API and AI Studio on the paid tier, in Gemini for business, and in Google Antigravity for paid Pro and Ultra subscribers. Developers use the model ID gemini-4-neon. There is no free tier in the API — if you want to try Neon without paying, the Gemini app on a Pro trial is the closest thing to a free test, and availability there depends on your plan.

For most readers, the two access points that matter are the app and the API. The app is the zero-setup way to try it: open Gemini as a Pro or Ultra subscriber and pick “Deeper analysis” in the model picker. The API and AI Studio are where developers will actually use it — building agents, wiring it into products, and running it at volume. AI Studio is the place to test the thinking levels and the long context window before committing to a production setup.

One availability note worth flagging: some users on telecom-bundled Pro plans have reported entitlement errors when trying to use Gemini 4 models inside Antigravity, according to launch-day coverage. If your Pro plan came bundled with a mobile carrier, check that Neon actually works in your account before assuming it does — and check the official site for the current list of supported plans.

Gemini 4 Neon vs the family — and the rivals

ModelBest forInput / 1MOutput / 1MIntelligence Index
Gemini 4 NeonHigh-frequency complex work$3.00$15.0061
Gemini 4 HydrogenFast everyday work$1.50$7.5051 / 54
Gemini 4 CarbonHardest long-horizon work$2 promo / $4 Dec 1$10 promo / $20 Dec 160
Claude Opus 5.5Anthropic's mid-tierCheck official siteCheck official site59
GPT-6.2 SolOpenAI's everyday model$2.00$10.0057
GPT-6.2 AstraOpenAI's flagship$8.00$40.0062
Claude Fable 5.2Anthropic's flagship$8.00$40.0065

Opus 5.5 pricing was not verified for this review — check the official site. Index scores from Artificial Analysis, November 18–19, 2026.

The table makes Neon's position clear. It is the only model in the top three that costs $3 per million input tokens: Fable 5.2 and Astra both cost $8, and Carbon costs $4 from December. Neon is one point behind Astra and four behind Fable on the index, but it costs less than half as much per token as either. Against its own siblings the value case is even sharper: one point ahead of Carbon on the index at 25% less than Carbon's standard price, and seven points ahead of Hydrogen at twice Hydrogen's price.

The comparison that matters most for buyers is Neon vs GPT-6.2 Sol, because Sol is the model most teams will actually cross-shop: same everyday-model role, $2 in and $10 out against Neon's $3 and $15, a million-token context on both sides, and an index gap of four points (61 vs 57). Neon costs 50% more per token and scores four points higher; whether that is worth it depends on whether your work hits the tasks where those four points show up — long-context reasoning, where Neon leads by 15 points on GraphWalks, and agentic coding, where it leads the composite. Our full Gemini 4 Neon vs GPT-6.2 Sol comparison covers that choice in detail. For the wider family picture, see our Gemini 4 Hydrogen vs Claude Fable 5.2 comparison and our best AI SEO tools guide for search and content work.

Strengths and limits

Gemini 4 Neon

Strengths

Weaknesses

Who Gemini 4 Neon is for — and who should skip it

Neon is for you if: you run coding agents every day and want the strongest agentic coding model on the composite index without paying flagship prices; you work with large codebases or long documents and want a million-token window with cheap cached input; you already pay for Google AI Pro or Ultra and want the “Deeper analysis” model in the Gemini app and Antigravity; or you build multi-step agent workflows where speed compounds and a 2x faster model means twice as many loops per hour.

Think twice if: your work is dominated by the hardest reasoning problems, where Carbon's deeper reasoning (or a flagship rival) earns its premium; if you need a free tier to experiment, because Neon has none — start with Hydrogen instead; if your budget is extremely tight, because Hydrogen does much of the same everyday work at half the price; or if you are choosing a model for a regulated or high-stakes use case where a one-day-old model with no independent audit history is a risk you do not need to take.

How to try it

The fastest path is the Gemini app as a Pro or Ultra subscriber: open gemini.google.com, pick “Deeper analysis” in the model picker, and start with a real task — a multi-file refactor, a long document summary, a research question with many sources. Developers should go to Google AI Studio, where the paid tier gives access with the model ID gemini-4-neon. Start with the default high thinking level, drop to medium or low for the trivial steps in your agent loops, and reserve xhigh for the steps where a mistake is expensive.

Before you commit, run your own test: take a real task from your own work, run it on Neon at high and on Hydrogen at xhigh, and compare the result and the bill. The benchmark tables will not tell you which model is right for your code; one afternoon of your own tasks will. And remember the December 1 price change — if you are comparing Neon with Carbon, do the maths at Carbon's $4/$20 standard rate, not the $2/$10 promotion that ends November 30.

What to watch in the coming weeks

Three things will decide whether Neon's launch-day story holds up. First, independent benchmarks: Terminal-Bench and the Artificial Analysis index are public methodologies, so third-party labs will rerun them within days, and their numbers — not Google's — will become the reference. Watch in particular whether the Coding Agent Index lead survives, and whether the shared-harness SWE-sweep figure moves.

Second, Carbon and Argon pricing from December 1: once the promotion ends, Neon becomes the clear mid-tier value pick — but Google could also change Neon pricing, and the whole family comparison moves with it. Third, real-world reliability in Antigravity: the planner-executor pattern is the feature that justifies Neon's existence, and developer reports over the next few weeks will show whether it works in messy, real codebases or only in Google's demos. We will update this review as independent results arrive.

Verdict: fast, strong, and fairly priced — with one asterisk

Gemini 4 Neon does what it promises, with one honest asterisk. It is fast — about twice Carbon's output speed — it is strong, ranking third in the world on the independent index and first on the coding-agent composite, and it is fairly priced at $3 in and $15 out, cheaper than Carbon's standard rate from December 1. The thinking levels give it a flexibility that single-setting models lack, and the Antigravity planner-executor role gives it a job no other Gemini model does at this price.

The asterisk is the SWE-sweep number: quote 39.0% only alongside the 12.8% shared-harness figure, and remember that Google's own comparison table is a selection, not a census. None of that changes the bottom line. For most developers, Neon is now the best all-rounder in the Gemini 4 family — the model to default to, with Hydrogen below it for volume work and Carbon above it for the hardest problems. Try it in AI Studio this week, put it on your real coding tasks, and let your own results decide.

Frequently asked questions

What is Gemini 4 Neon?

Gemini 4 Neon is Google's mid-tier AI model, launched on November 18, 2026. Adapted from the flagship Gemini 4 Carbon, it is built for complex coding, analytical reasoning, enterprise knowledge work and science, with a 1,048,576-token context window and four thinking levels. It sits between the fast, cheap Hydrogen and the flagship Carbon in the Gemini 4 family.

How much does Gemini 4 Neon cost?

Google charges $3 per million input tokens and $15 per million output tokens on the standard paid tier of the Gemini API, with cached input at $0.15 per million tokens. Batch and Flex tiers cost half the standard rate. There is no free tier. Prices are from launch week, November 2026 — check the official site for current pricing. Note that Carbon's $2/$10 promotional pricing ends November 30, 2026, after which Carbon costs $4/$20.

Is Gemini 4 Neon better than Claude Opus 5.5?

On the independent Artificial Analysis Intelligence Index, Neon scores 61 against 59 for Opus 5.5, and it leads the Coding Agent Index at 58.6. But Opus edges Neon on two individual coding benchmarks: Terminal-Bench 4.1 (65.8% vs 65.2%) and FrontierCode 2.0. Google's own comparison table shows Neon ahead on most of its 18 published benchmarks, but that is Google's selection — independent checks show Neon winning about half the rows outright. The fair summary is that Neon is stronger overall on the composite measures, while Opus still wins a few individual tests.

What are the thinking levels in Gemini 4 Neon?

Neon has four thinking levels: low, medium, high (the default), and xhigh. Higher levels mean more internal reasoning before answering — slower and more expensive, but usually more accurate. The level moves the benchmark scores: Neon scores 60 on the Artificial Analysis Intelligence Index at high and 61 at xhigh. Use high by default, low or medium for simple tasks, and xhigh where accuracy is worth waiting for.

What is the SWE-sweep controversy about Gemini 4 Neon?

Google reports Neon at 39.0% on SWE-sweep, but that figure was measured inside Google Antigravity, which runs parallel subagents across a codebase. In the shared mini-SWE harness that all models can use, the same model scores 12.8%. Opus shows the same pattern: 10.1% in Claude Code against 3.6% in the shared harness. The headline number compares agent products as much as models — on the shared harness Neon still leads, but by about one point over GPT-6.2 Sol, not 29.

Where can I use Gemini 4 Neon?

Neon is available in the Gemini app for AI Pro and AI Ultra subscribers, through the Gemini API and AI Studio on the paid tier, in Gemini for business, and in Google Antigravity for paid Pro and Ultra subscribers. Developers use the model ID gemini-4-neon. There is no free tier in the API. In Antigravity, Neon can act as an Orchestrator, planning work and delegating tasks to cheaper Hydrogen subagents.

AT
About Aitopkit

We research AI tools in depth before recommending them — official pricing, public docs, and real user reports. If you buy through our links we may earn a commission at no extra cost to you. Read our disclosure.