AI·News & analysis
Google's Gemini 4 Neon scores 61 and costs $3 per million tokens
Google released Gemini 4 Neon, a model that aims for top-tier reasoning at a faster pace. It costs $3 per million input tokens and is live in the Gemini API and the Gemini app for Pro and Ultra.

Tide
Wave
Sci-fi
5/10
Reality
Shipping
A mid-sized brain that thinks like a big one.How we rate
Google released Gemini 4 Neon, a faster AI model built for coding, office work and science tasks.
It scores 61 on the Artificial Analysis Intelligence Index, one point above Gemini 4 Carbon and third overall. The API costs $3 per million input tokens and $15 per million output tokens. Neon is live for developers today and in the Gemini app for Pro and Ultra subscribers, with a security version coming to trusted defenders.
What to know
- Google released Gemini 4 Neon on November 18, the fourth Gemini 4 model in about seven weeks, after Argon, Carbon and Hydrogen.
- Neon scores 61 on the Artificial Analysis Intelligence Index at its top thinking level, one point above Carbon and third overall.
- The API costs $3 per million input tokens and $15 per million output tokens, with a 1 million token context window.
- It is live in the Gemini API, Google AI Studio and Antigravity, and Pro and Ultra subscribers can pick it in the Gemini app.
Google announced Gemini 4 Neon on November 18, a new AI model it says brings top-tier reasoning with fast, end-to-end task completion. It is the fourth Gemini 4 model in about seven weeks.
So what does Neon do that Carbon, Argon and Hydrogen do not? Let's start with where it fits.
Where does Neon fit?
The Gemini 4 family now has four models, each with a job.
- Carbon and Argon: the big frontier models, built for deep, long reasoning.
- Hydrogen: the fast one for everyday work, launched on November 12. See our story on Hydrogen.
- Neon: the middle path. Google says it balances depth and efficiency for complex jobs that come up again and again.
Google describes Neon as built for complex coding, analytical reasoning and production-grade work at companies. In plain words, it is meant to be strong enough for hard tasks and quick enough to run all day.
How does it score?
Google points to the Artificial Analysis Intelligence Index, a score from an outside company that tests AI models. At its top "xhigh" thinking level, Neon scores 61. That is third overall, just behind larger and slower models.
The same index puts Carbon at 60, so Neon is the highest-scoring Google model the index has measured. GPT-6.2 Astra scores 62 and Claude Fable 5.2 scores 65 on that chart.
By the numbers: Google lists many more results, mostly from outside testers.
- Vals Index: second place, at 70.9%. A benchmark is a standard test used to compare models.
- Agent Arena Index: second place with 1336 points, across more than 4,200 real agent sessions. An AI agent is a program that takes several steps on its own.
- Coding Agent Index: first place, at 58.6, according to Google's post.
- Terminal-Bench 4.1: 65.2%, a test of how well a model handles tasks in a command-line terminal.
- Science: 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond.
These are Google's reported figures. Independent testers will keep running their own checks over the coming days.
What does it cost?
For developers, the price is simple. An API is a doorway that lets apps plug into a model. Using it is billed in tokens, small chunks of text of about four characters each.
- Input: $3 per million tokens.
- Output: $15 per million tokens.
- Cached input: $0.15 per million tokens, a 95% discount on text the model has already seen.
A context window is how much text the model can read at once. Neon's is 1 million tokens, and it can write up to 1 million tokens in one answer. It takes text, images, video, audio and PDFs, and it gives text back. It does not generate images or audio.
Here is the pricing wrinkle. Carbon and Argon are on a launch discount of $2 and $10 per million tokens until the end of November. From December 1 they go up to $4 and $20. Neon's $3 and $15 is a quarter below that standard price, but 50% above the discounted one. Hydrogen costs half as much, at $1.50 and $7.50.
Is it really faster and cheaper?
According to Artificial Analysis, mostly yes. Running its whole index with Neon at xhigh costs $6.44, about 17% less than the $7.76 for Carbon at its standard price. Neon uses about 15% more output tokens per task, so the saving comes from the lower price.
At the default "high" level, Neon matches Carbon's score of 60 for $3.74 per task, less than half of Carbon's cost. Speed is the other gain. Artificial Analysis measures about 110 tokens per second for Neon against about 57 for Carbon, which cuts the time per task by roughly 40%.
Google adds its own numbers. At the high level, Neon costs 32% less per task than Carbon. It says that is roughly 50% to 80% less per task than top rival models. Even at xhigh, Google says it costs more than 36% less per task than rivals at their highest setting.
OpenAI's GPT-6.2 Astra is still the cheaper option at the top. It finishes the index for about $4.00 per task, so Neon at xhigh costs roughly 1.6 times as much for a score one point lower.
What about coding?
This is where Google puts most of its effort. Neon is built for long jobs that use many tools, and Google says it cuts down on reading the same file twice and handles long sessions well.
Inside Google Antigravity, Google's coding tool, Neon can act as an orchestrator. That is a lead model that hands work to helper agents, called subagents. In this case the helpers are Hydrogen models running in parallel.
In real life Say you ask an assistant to fix a bug that touches ten files. The lead model plans the fix and splits it up. Fast helpers edit the files and run the tests, and the lead checks the result.
Artificial Analysis found Neon scores 59 on its Coding Agent Index when run with Google's Antigravity command-line tool, three points above Carbon. An average task costs $6.86 with Neon, against $12.61 for Claude Opus 5.5 and $15.47 for Claude Fable 5.2 on the same test.
One more number shows what the tool around a model can do. On a benchmark called SWE-sweep, Neon scores 12.8% in a plain test setup. In Antigravity it more than triples to 39.0%.
Google also shared demos built with Neon:
- A ski game: a first-person WebGL2 ski simulator on real Mont Blanc terrain, with 10,200 lines of shader code and no ready-made assets.
- A 3D study: Alan Turing's 1936 study at King's College, Cambridge, rebuilt in interactive 3D, with a working Universal Turing Machine on the desk.
- A spreadsheet picture: one image turned into a 100,000-cell Google Sheet using more than 500,000 live formulas.
What can it do for office work?
Most people will meet Neon through work tasks rather than code. Google says it reads spreadsheets, PDFs and photos more precisely and turns dense material into reports and summaries. It can also pull details from Gmail and Docs.
It tops the Legal Research Bench and the Tax Agent Bench, according to Google. Artificial Analysis says Neon's biggest gains are in knowledge work and visual reasoning. On its AA-Briefcase test, Neon scores 1700 Elo, 89 points above Carbon. An Elo score ranks models by head-to-head wins, like chess ratings. The biggest jump is in slide quality, up 170 points to 1771, the highest that testers have measured.
In the Gemini app, tasks can now run on Neon. Google gives examples like checking billing statements to find out why a utility charge spiked, or cross-checking receipts to argue for a refund. Google says this stretches your usage limits further than handing the same jobs to a bigger model.
Who gets the security version?
Google is also releasing Neon for cyber. It is aimed at defenders who find and fix flaws in code. Eligible members of the Fairwind Program start getting it over the next few days, and others can ask their Google Cloud account team.
Google's numbers cover the whole repair cycle:
- Finding flaws: 90.5% on SEC-bench Pro, at roughly half the cost per challenge.
- Checking risk: on ExploitGym it matches frontier results with half the tokens of Carbon.
- Fixing flaws: 70.7% on CWE-bench v2, up 3.4% from Carbon.
- Safer code: 76% on the hard tier of AutoBaxBench, which Google calls a new best.
We covered the defenders-only launch of its older sibling in Gemini 4 Argon for cyber defenders.
How safe is it?
Google says Neon keeps Carbon-level protections against severe misuse, such as help with cyberattacks or biological threats. It also says it refuses less legitimate work, declining just 1.98% of tasks on Terminal-Bench 4.0 and 0% on the science set.
The company also reports a lower rate of cheating on coding tests than any other Gemini 4 model. It reports a 0.09% misaligned action rate on its agent safety suite and a 99.2% pass rate on an honesty test. On its behavior tests it cut overreach, where an agent does more than you asked, by 15% to 35%.
Then there is prompt injection. That is when hidden instructions in a web page or email try to trick an AI. With its safeguards on at the top thinking level, Neon resisted every attack in the Gray Swan test. These are Google's own tests, so treat them as claims until outsiders repeat them.
Where can you get it?
Neon is rolling out today across several places.
- Developers: the paid tier of the Gemini API in Google AI Studio and Android Studio.
- Subscribers: Google AI Pro and Ultra members can use it in Antigravity, and all Pro and Ultra members can use it in AI Studio.
- Businesses: Gemini for business, through the API or the model drop-down, and Antigravity.
- Gemini app: pick it from the model drop-down if you pay for AI Pro or Ultra.
Free users do not get Neon for now. Our story on Gemini 4 Argon covers how the first model in this family reached the app.
What it means for you
- Developers: Neon is a cheaper, faster option than Carbon at the standard price, and it works well inside Antigravity.
- Teams: check the December 1 price change on Carbon and Argon before you choose a model for a big job.
- Gemini subscribers: try Neon in the drop-down for long research and document jobs, and keep the default for quick questions.
- Everyone else: nothing changes yet, but this is the kind of model that will reach free tiers later.
The bottom line
Google now has four Gemini 4 models in seven weeks, and Neon is the one that aims at the sweet spot of strong and fast. Outside tests put it one point above Carbon at a lower cost, though rivals still lead in places. The big claims are Google's, so watch the independent results this week.
Key facts
- API price
- $3 input, $15 output per million tokens
- Context window
- 1 million tokens
- Intelligence Index
- 61 at xhigh, Carbon scores 60
- Speed
- About 110 tokens per second vs 57 for Carbon
- Gemini app access
- Google AI Pro and Ultra subscribers
Got questions?
Quick answers, plain wordsWhat is Gemini 4 Neon?
It is Google's newest AI model, released on November 18, 2026. Google says it brings frontier-level intelligence, meaning top-tier reasoning, with fast end-to-end task completion for software engineering, knowledge work and science.
How much does Gemini 4 Neon cost?
Developers pay $3 per million input tokens and $15 per million output tokens. Cached input, which means text the model has already seen, costs $0.15 per million tokens, a 95% discount. A token is a small chunk of text, about four characters.
Is Gemini 4 Neon free?
No. In the Gemini app it is limited to paid Google AI Pro and Ultra subscribers. Developers use the paid tier of the Gemini API in Google AI Studio and Android Studio, and Pro and Ultra subscribers can also try it in AI Studio.
How is Neon different from Carbon and Hydrogen?
Google says Carbon is for your hardest, longest projects and Hydrogen handles fast everyday work. Neon sits between them and balances depth and efficiency for complex, high-volume workflows.
Is Neon smarter than Carbon?
Slightly, on one outside test. At its xhigh thinking level Neon scores 61 on the Artificial Analysis Intelligence Index and Carbon scores 60. Neon is also faster and costs less per task at the standard price.
What are the thinking levels?
Neon has four levels, low, medium, high and xhigh. Higher levels reason longer, which costs more and takes more time. High is the default, and Google says each level scales steadily so teams can pick what a job needs.
How big is the context window?
One million tokens. A context window is how much text the model can read at once. Google says Neon can combine spreadsheets, PDFs, charts, audio and video in that single window.
What is Neon for cyber?
It is a version of Neon tuned for security work, such as finding, checking and fixing code flaws. It is going first to eligible members of Google's Fairwind Program, starting over the next few days.
Is Neon safe to use with web pages and emails?
Google says it resisted every attack on a Gray Swan test that plants hidden instructions in web pages, emails and documents, when its real-time safeguards were on at the highest thinking level. That is Google's own report.
Where can I try Gemini 4 Neon today?
Pick it from the model drop-down in the Gemini app if you have Google AI Pro or Ultra. Developers can use the Gemini API, Google AI Studio and Google Antigravity, and businesses get it in Gemini for business.
SourcesGoogle
Topics and tagsGoogle, AI agents, google, gemini
Go deeper
Earlier on this story
Related guide
Guide, 8 minHow to choose an AI assistant: ChatGPT, Claude, Gemini and moreA plain-English way to pick between ChatGPT, Claude, Gemini, Copilot, Perplexity and Meta AI: start free, match the tool to your main...Related stories

Google's Gemini 4 Hydrogen beats pricier Argon on one key score
Google released Gemini 4 Hydrogen, a fast model for everyday coding and agent work. It costs $1.50 per million input tokens and reaches the Gemini app and Search today.

Google makes Gemini Tasks free in six countries, with calls and payments
Google's Gemini app can now take on errands for free in the US, Japan, Korea, Canada, Australia and India. In the US it can also phone businesses and pay with Google Pay.

Four Gemini 4 Argon agents beat Minecraft with no human help
Google DeepMind researchers gave four AI agents one goal, defeat the Ender Dragon, and no game tools. Twenty-seven hours and 605 deaths later, the dragon fell.
More in brief
- ASELSAN's EJDERHA electromagnetic weapon disables 51 drones with one commandNov 21
- Microsoft To Do will shut down on June 30, 2027Nov 21
- Medtronic's Hugo surgical robot cleared for general surgery in the USNov 21
- Google's Gemini now turns noisy fire radio into text for Watch DutyNov 21
- Internet Archive staff move to unionize with CWANov 20
- Etsy sellers are leaving over new terms that allow AI trainingNov 20