Gemini 4 Neon (2026): Price, Benchmarks, Who Gets It and Where It Fits Between Hydrogen and Carbon

Jump to comments

Five glass gas-discharge tubes labeled He, Ne, Ar, Kr and Xe glowing orange, red, violet, pale blue and white against a black background
Discharge tubes of helium, neon, argon, krypton and xenon, each glowing its own color under high voltage.
Photo: Alchemist-hp, 2009, via Wikimedia Commons (CC BY-SA 2.0 DE)

Google released its fourth Gemini 4 model in seven weeks on Wednesday, Nov. 18, 2026. Gemini 4 Neon is a middle model. Google pitches it as frontier-level intelligence at higher speed and lower cost, and the first outside tests put it level with Gemini 4 Carbon, Google’s strongest model. It went live in the Gemini API, Google AI Studio, Gemini for business and the coding app Google Antigravity, and in the Gemini app for paying Google AI Pro and Ultra subscribers. Within hours "Gemini 4 Neon" was one of the top trending topics on X in the United States. On the same afternoon, Google DeepMind opened its Genie 4 world model to some AI Ultra subscribers.

The short answer: Neon costs $3 per million input tokens and $15 per million output tokens. That is twice the price of Gemini 4 Hydrogen and 25% below Carbon’s standard price, which starts on Dec. 1. It keeps the family’s 1 million-token context window and four thinking levels. The independent firm Artificial Analysis scores it 61 on its Intelligence Index, third overall and one point ahead of Carbon. Its main gains are speed and cost per task, not a big jump in capability. If you pay for Google AI Pro, Neon is the strongest Gemini model you can now pick in the app. Free users get no new access.

What Google released on Nov. 18

The announcement came from Koray Kavukcuoglu, who leads model development at Google DeepMind and is Google’s chief AI architect. Google pitches Neon at the work that sits "right in the middle," as its Antigravity team put it: coding and office jobs complex enough to need careful reasoning, but frequent enough that waiting minutes for each answer gets old. The model handles text, images, audio and video as input and writes text as output.

The details that matter most:

Price. $3 per million input tokens and $15 per million output tokens. Cached input, meaning repeated prompt text the system has already seen, costs $0.15 per million tokens, a 95% discount. Neon is on the paid API tier only; there is no free tier, unlike Hydrogen.

Context. Up to 1,048,576 tokens in one request, enough for several long books or a large code repository, according to Google Cloud’s model page. Answers can also run very long, up to the same limit when developers use background mode.

Thinking levels. Low, medium, high and extra high ("xhigh"). High is the default. More thinking means better answers on hard problems but more tokens billed and more waiting.

Lineage. Google’s model card says Neon is "adapted from Gemini 4 Carbon" and points to Carbon’s model card for its training data, hardware and software. Its knowledge cutoff is March 2026.

Where it runs. The Gemini API paid tier in AI Studio and Android Studio; Gemini for business and Google Cloud’s agent platform; Google Antigravity for paid AI Pro and Ultra subscribers, through the desktop app, command-line tool and editor extensions; and the model menu of the Gemini app for Pro and Ultra subscribers. A security-focused "Neon for cyber" goes to approved defenders in Google’s Fairwind Program over the next few days.

9to5Google reports that in the Gemini app, Neon replaces both Gemini 4 Argon and the older Gemini 3.1 Pro in the model picker for paid users. For people on AI Pro, that is the real upgrade: Carbon in the app has so far been limited to the more expensive Ultra plan.

Where Neon fits in the Gemini 4 family

All four Gemini 4 models are named after chemical elements, and Google’s launch demos leaned into the joke. In one, Neon built a 3D lab bench with Bunsen flames for Carbon and Hydrogen and glowing gas-discharge tubes for Argon and Neon. Here is the lineup in release order:

Gemini 4 Argon (announced Sept. 30, widely available Oct. 13)

The first Gemini 4 model, and the one that ended what Reuters called months of delays for Google’s flagship line. Argon went to security testers first and then to developers, businesses and Pro and Ultra subscribers on Oct. 13. The Gemini app described it as the "ideal choice for the majority of software engineering and knowledge work." It launched at an introductory $2 per million input tokens and $10 per million output tokens.

Gemini 4 Carbon (announced Oct. 28, widely available Nov. 10)

Google’s strongest model, built for the longest and hardest jobs, from deep debugging to scientific proofs and security audits. Google’s developer guide says Carbon "spends more thinking tokens per turn to explore thoroughly and verify its work," which makes it slower. Carbon costs the same introductory $2 and $10 through Nov. 30, then $4 and $20 from Dec. 1. In consumer apps it is for Ultra subscribers.

Gemini 4 Hydrogen (Nov. 12)

The fast, cheap model for everyday coding, apps and running many small agent tasks in parallel. It costs $1.50 and $7.50 per million tokens and is the only Gemini 4 model with a free API tier. It also powers Google Search and the Gemini app for free users when the app picks it automatically. Google says it scored 51 on the Artificial Analysis index at its default setting, one point ahead of Argon’s 50, and 54 at extra high.

Gemini 4 Neon (Nov. 18)

The new middle: Carbon-class answers at what 9to5Google describes as "Flash speeds." In Antigravity, a new Orchestrator mode pairs the two newest models in one session. Neon plans the work and Hydrogen carries out the routine edits, package updates and test runs. When Hydrogen gets stuck, it hands control back to Neon.

While Carbon tackles your hardest, long-horizon projects and Hydrogen handles fast everyday execution, Neon balances depth and efficiency. (Koray Kavukcuoglu, Google DeepMind, Nov. 18, 2026)

Prices were checked on November 18, 2026, on Google’s own pages. Carbon’s step up to $4 and $20 on Dec. 1 is what makes Neon look cheap: before then, Neon actually costs 50% more per token than Carbon’s promotional rate.

What Google claims

Google’s announcement is a long list of first and second places. The headline numbers, all from Google’s blog post and model card:

Overall intelligence

Third overall with a score of 61 on the Artificial Analysis Intelligence Index at extra high, behind only "larger, slower models." Second on the Vals Index at 70.9%, and second on Arena’s new Agent Arena Index with a score of 1336 across more than 4,200 real agent sessions.

Coding

First on the Artificial Analysis Coding Agent Index at 58.6. Google also reports 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at the default setting. On SWE-sweep, a test of finding and fixing bugs without breaking working code, Neon scores 12.8% in a minimal test harness and 39.0% inside Antigravity. Google says that shows what its own agent tooling adds.

Office and business work

First on Zapier’s AutomationBench at 56.3%, and first on Andon Labs’ Vending-Bench 2, a long simulation of running a small business, finishing with $16,843. Google also reports 75.8% on Agents’ Last Exam v1.1 and 73.5% on Chartography, a test of reading dense professional charts.

Science and math

Best in its class on Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%).

Safety

Google says Neon has Carbon-level safeguards against cyber and biological misuse while refusing far less legitimate work. It turned down 1.98% of Terminal-Bench 4.0 tasks. It took shortcuts on 2.45% of tasks in Google’s coding cheating test, the lowest of any Gemini 4 model, and passed 99.2% of Google’s agent honesty checks. With real-time safeguards at its highest thinking level, it resisted every attack in Gray Swan’s indirect prompt injection benchmark. That benchmark plants hidden instructions in web pages, emails and documents.

What independent testing says so far

Most of those numbers come from Google or from leaderboards Google chose to cite. The most complete outside test so far is from Artificial Analysis, an independent benchmarking firm that publishes its methods. Its verdict is in the headline it used: "a small Intelligence gain over Gemini 4 Carbon at a ~17% lower Cost per Task."

Capability is about the same as Carbon. Neon at extra high scores 61, one point above Carbon at high. Artificial Analysis ranks it behind OpenAI’s GPT-6.2 Astra and Anthropic’s Claude Fable 5.2, each at its maximum setting. The firm found small improvements in agentic office work, visual reasoning and coding.

It is cheaper because the tokens are cheaper. Each Artificial Analysis task cost $6.44 on Neon against $7.76 on Carbon at Carbon’s standard price. Neon actually used about 15% more output tokens per task. The saving comes entirely from its 25% lower token prices.

It is much faster. Neon averaged about 21 minutes per task against about 36 for Carbon, because it writes about 110 output tokens per second against about 57.

It is a bargain as a coding agent. Running in the Antigravity command-line tool, Neon scored 59 on the firm’s Coding Agent Index, three points above Carbon. That is level with Claude Code running Claude Fable 5.2 (58) or Claude Opus 5.5 (58). The cost is the striking part: about $6.86 per task for Neon, against $12.61 for Opus 5.5 and $15.47 for Fable 5.2.

Google’s own model card also shows where Neon does not lead. Anthropic’s Claude Opus 5.5 edges it on Terminal-Bench 4.1 (65.8% to 65.2%) and FrontierCode 2.0 (47.3% to 46.2%). OpenAI’s GPT-6.2 Sol beats it on Chartography (75.1% to 73.5%). For more on OpenAI’s newest models, see our explainer on GPT-6.2 Astra and Sol.

Two cautions. Some launch-day summaries circulating on X put Neon’s Artificial Analysis score at 67. The firm’s own published figure is 61. And launch-day benchmarks come from short tests. How Neon holds up over weeks of real coding and office work, the thing developers care about most, will only show over time.

Glass entrance of a multi-story office building with a lit Google logo above revolving doors and a cyclist outside
The Google building at 6 Pancras Square in London’s King’s Cross, where Google DeepMind is based.
Photo: Gciriani, 2017, via Wikimedia Commons (CC BY-SA 4.0)

Who can use Neon, and what it costs you

Free Gemini users. Nothing changes today. Free users still get Hydrogen when the app picks it automatically, and Hydrogen now also powers Google Search for everyone.

Google AI Pro subscribers. Choose Gemini 4 Neon from the model menu in the Gemini app. It is the biggest step up for this tier, because Carbon in the app has been reserved for Ultra. Pro subscribers can also use Neon in Antigravity and AI Studio.

Google AI Ultra subscribers. You keep Carbon and gain Neon as a faster option. Ultra is also the only plan that gets the new Project Genie with Genie 4, below.

Developers. Call the model ID gemini-4-neon on the paid API tier. Google’s guidance is simple: Hydrogen for fast, high-volume work and parallel subagents; Neon for most serious coding and analysis; Carbon for the hardest, longest problems, where its extra thinking is worth the wait and the bill. Every token is billed, including the model’s thinking tokens and the hidden instructions that come with search grounding. Higher thinking levels cost real money on long jobs.

Businesses. Neon is in Gemini for business and Google Cloud’s agent platform, with batch and provisioned-throughput options.

If you mostly use AI through tools that plug into many models, such as cloud AI agent teams or apps connected through the Model Context Protocol, Neon will usually appear as one more model choice. Google’s own demos ran it through a Blender connector to build 3D scenes.

The same-day launch: Genie 4

Neon was not Google’s only release on Wednesday. Google DeepMind and Google Labs opened Genie 4, which Google calls its most realistic and interactive world model yet, to eligible Google AI Ultra subscribers in Project Genie, a browser-based research prototype. AI Pro members can join a waitlist.

A world model generates a place you can move through in real time, like a video game that is drawn as you play. With Genie 4 you can type a new prompt mid-session to change the weather, add a character or redirect the action. For the first time, it also generates matching sound as the scene unfolds, from ambient noise to character dialogue. Google’s model card says it streams 720p video at up to 24 frames per second and is built on Gemini Omni Flash.

Google DeepMind’s introduction to the new Project Genie, powered by Genie 4, released Nov. 18, 2026.

The update also lets Genie 3 run up to three minutes per generation, up from 60 seconds. It adds an agent that builds worlds from a conversation and a community gallery for sharing and remixing worlds. Google says both models are still research systems: details can drift in long sessions, and readable text and complex physics remain hard. Everything made in Project Genie carries Google’s invisible SynthID watermark.

Should you switch?

Stay on Hydrogen if your work is quick and high volume, like chat features, document summaries or many small agent steps, or if you need the free tier. Neon costs twice as much per token.

Try Neon if you use Carbon today and find it slow, or if you are on AI Pro and want Google’s best reasoning in the Gemini app. On Artificial Analysis’s numbers, you lose almost nothing in capability and finish tasks in a little over half the time.

Keep Carbon for the hardest problems where an extra point of accuracy matters more than time or money. Plan for its price to double on Dec. 1.

Compare before committing if you already pay for Claude or OpenAI models. On outside tests, Neon trades blows with Opus 5.5 and Fable 5.2 on coding at a much lower cost per task, but it is not the outright leader.

What is still unclear

Usage limits. Google has not published how many Neon prompts Pro subscribers get per day in the Gemini app, only that tasks on Neon stretch usage limits further than larger models.

Argon’s future. Neon replaces Argon in the Gemini app, but Google has not said how long Argon stays in the API.

Real-world reliability. Benchmarks and launch demos, like a ski simulator written in 10,200 lines of shader code, show what is possible. They do not show how often the model gets ordinary work right. The first weeks of developer feedback will matter more than the launch-day leaderboards.

The pace. Four Gemini 4 models in seven weeks is unusually fast, even for this industry. Artificial Analysis titled its review "Google continues to accelerate its release cadence." Whether Google keeps that pace into December is the next thing to watch.

The bottom line

Gemini 4 Neon is not a leap in raw intelligence. It is Google packaging close to Carbon-level answers into something faster and, from December, cheaper, and handing it to the much larger group of AI Pro subscribers. For developers it fills the gap between cheap Hydrogen and slow, costly Carbon. For everyone else, it is the model to try in the Gemini app this week.

Further reading

Google Pixel Buds 2a wireless earbuds in Iris with their charging case
Google Pixel Buds 2a

Google’s noise-cancelling earbuds, listed with built-in help from Gemini, the company’s AI assistant.

Logitech MX Keys S low-profile wireless keyboard in graphite
Logitech MX Keys S Wireless Keyboard

A quiet, low-profile wireless keyboard that switches between up to three computers, a common pick for long days in AI coding tools.

Logitech MX Anywhere 3S compact wireless mouse in graphite
Logitech MX Anywhere 3S Compact Wireless Mouse

A compact wireless mouse with a fast-scrolling wheel for long documents and code files, and quiet clicks.

Cover of Supremacy by Parmy Olson
Supremacy: AI, ChatGPT, and the Race That Will Change the World

Parmy Olson’s account of the race between Google DeepMind’s Demis Hassabis and OpenAI’s Sam Altman, the Financial Times Business Book of the Year.

Cover of Genius Makers by Cade Metz
Genius Makers: The Mavericks Who Brought AI to Google, Facebook, and the World

Cade Metz’s history of the researchers who brought deep learning to Google and its rivals, including the founding of DeepMind.

Aglena Best Of product picks

Leave a Comment

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

Scroll to Top
Aglena