Google Gemini 4 Neon: Faster Frontier AI at $3 Per Million

Published (Updated )

Quick answer
- Google launched Gemini 4 Neon on November 18, 2026, a new AI model built for complex, high-frequency work such as coding, analysis and science.
- It scores 61 on the Artificial Analysis Intelligence Index at its highest thinking level, ranking third overall, just behind larger, slower frontier models.
- It costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million tokens, and it has a 1 million-token context window.
- It is rolling out now in the Gemini API, Google AI Studio, Gemini for business, Google Antigravity and the Gemini app for paid Google AI Pro and Ultra subscribers. It is not available on the free plan.
- Neon sits between Gemini 4 Carbon, built for the hardest long-horizon projects, and Gemini 4 Hydrogen, built for fast everyday tasks.
What Gemini 4 Neon is
Google is releasing AI models at a pace that would have seemed absurd a year ago. Gemini 4 Neon is the fourth model in the Gemini 4 family in just seven weeks, following Carbon and Hydrogen the week before. Each one is aimed at a different slice of work, and Neon is the one for jobs that are both hard and frequent: production code, multi-step analysis, enterprise documents and scientific research, done fast enough to run all day.
Google's own description is that Neon brings "frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows." In plainer terms, it is meant to be the model you reach for when a task is too demanding for a lightweight model but does not need the biggest, slowest model in the lineup, and when you need that answer quickly and cheaply enough to do it a hundred times a day.
The model was announced on November 18, 2026, on the Google blog, and it began rolling out the same day. It is available through the Gemini API paid tier in Google AI Studio, in Gemini for business through the API and model dropdown, in Google Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. It is not available on the free Gemini plan.
Where Neon sits in the Gemini 4 family
Google now has three named models in the Gemini 4 range, and the differences between them are about what kind of work each one is tuned for rather than a simple ranking from worst to best.
- Gemini 4 Carbon is the flagship for the hardest, longest projects. It takes on the work that needs the deepest reasoning over the longest time horizons, and it is the slowest and most expensive of the three.
- Gemini 4 Neon is the middle model. It is built for complex coding, analytical reasoning and production-grade enterprise workflows, and it is designed to handle those jobs at high frequency, meaning many tasks in a row, quickly and at a lower cost.
- Gemini 4 Hydrogen is the fast everyday model, aimed at quick tasks where speed matters more than maximum depth.
Neon replaces Argon in the lineup, and the older Gemini 3.1 Pro is being retired for paid users, with the 3.5 Flash-Lite model now appearing as "Auto" in the model picker. The practical effect is that paid users get a cleaner choice: Hydrogen for quick work, Neon for serious work done often, Carbon for the hardest work.
Neon also has four thinking levels, Low, Medium, High and Extra High, so teams can dial how much reasoning, and how many tokens, each task gets. Its highest intelligence scores come at the Extra High level, while its default High level is where Google quotes most of its cost comparisons.
How Neon scores on benchmarks
Benchmark numbers need care. Some of these are independent evaluations, some are Google's own tests, and a model that tops one list can trail on another. With that in mind, Neon's headline results are strong, and they are consistent across several independent indexes.
- Artificial Analysis Intelligence Index: 61 points, ranked third overall. At its Extra High thinking level, Neon sits just behind larger, slower models. Artificial Analysis also notes it delivers a small intelligence gain over Carbon at roughly 17 percent lower cost per task.
- Vals Index: ranked second at 70.9 percent, and first on both the Legal Research Bench and the Tax Agent Bench. The Vals Index measures real-world professional work rather than textbook questions.
- Artificial Analysis Coding Agent Index: ranked first overall at 58.6. This is the index that matters most for developers, because it measures how well a model works as a coding agent, not just how well it writes a single function.
- Agent Arena Index: ranked second globally with a score of 1336, across more than 4,200 real-world agent sessions. Neon also ranks second on steerability, which measures how cleanly a model adapts when a user corrects it mid-task.
- Terminal-Bench 4.1: 65.2 percent, and FrontierCode 2.0: 46.2 percent at its default High setting. On SWE-sweep, it leads the mini harness at 12.8 percent and more than triples to 39.0 percent inside Google Antigravity.
- Science benchmarks: Neon is best in class on Terminal-Bench Science at 70.5 percent, RiemannBench at 84.0 percent and HLE-Diamond at 71.3 percent.
- Enterprise and automation: It ranks first on most Surge AI benchmarks across finance, accounting, spreadsheet modelling and complex constraint following, scores 73.5 percent on Chartography for reading dense professional charts, and ranks first overall on Zapier's AutomationBench at 56.3 percent and on Andon Labs' Vending-Bench 2, a long-horizon simulation of running a business, with a score of $16,843.
- Agents' Last Exam v1.1: 75.8 percent, on par with leading models.
Taken together, the picture is consistent: Neon is not the single smartest model Google makes, but it is close to the top on most measures, and it is the best in its class at coding-agent work, which is exactly the job Google built it for.
What it costs
Pricing is where Neon makes its case most directly, because the model is aimed at work that happens hundreds of times a day, where small per-token differences add up fast.
- Input tokens: $3 per million, with cached input at $0.15 per million, a 95 percent discount on repeated context.
- Output tokens: $15 per million.
- Context window: 1 million tokens for input and output combined, matching the rest of the Gemini 4 family.
- Cost per task: 32 percent lower than Carbon at Neon's default High thinking level, for a similar Artificial Analysis Intelligence Index score. On the Vals Index and Agent Arena, Neon sits on the Pareto frontier, the set of models where no rival is both smarter and cheaper, at roughly 50 to 80 percent lower cost per task than top competing frontier models. At Extra High thinking, it still costs 36 percent or more less per task than competing models at their highest setting.
For comparison, Argon and Carbon are priced at $2 per million input tokens and $10 per million output tokens until November 30, after which they rise to $4 and $20. Neon is therefore 25 percent cheaper than Carbon and Argon will be once that promotional pricing ends, according to 9to5Google.
For a US business running coding agents all day, that pricing is the point. A model that is nearly as capable as the flagship but costs a third less per task changes how many tasks you can afford to automate, and how often you can afford to let an agent retry, check its own work or run a longer job.

Built for production code
Google's pitch for Neon is not just about benchmark scores. It is about how the model behaves in long, tool-heavy jobs, which is where coding agents usually fall apart.
Neon is engineered for long-running agent workflows, and it can act as a fast, high-judgment orchestrator that directs parallel Hydrogen subagents inside Google Antigravity. Google says it cuts down on redundant file reads, shares progress as it works, and handles context compaction reliably in long sessions, the kind of session where a model has to keep track of a large codebase over hours.
For production-ready code, Google says Neon handles multi-file debugging and refactoring across repositories with targeted, surgical edits, rather than rewriting whole files. That is a claim about behaviour rather than raw intelligence, and it is the kind of thing that matters to a team deciding whether to trust a model with a real codebase.
The model also reads across formats in a single context window. Enterprise work lives in spreadsheets, PDFs, charts, audio and video, and Neon can synthesise across all of them and turn the result into structured deliverables. In the Gemini app, it writes with well-structured responses and a natural, direct tone, reads uploads like spreadsheets, PDFs and photos more precisely, and pulls in details from Google apps like Gmail and Docs.

A version for cyber defence
Alongside the main model, Google is releasing Gemini 4 Neon for cyber, a variant tuned for defensive security work. It is rolling out to trusted defenders over the coming days as part of the Fairwind Program, Google's programme for giving security teams early access to frontier models.
The numbers Google quotes for the cyber variant are striking:
- AutoBaxBench: Neon produces more secure code than Carbon across easy, medium and hard challenges, and reaches state of the art on the hard tier at 76 percent.
- WebExploitBench v1.2: state of the art at 51.26 percent, at half the cost of competing large models.
- SEC-bench Pro: state of the art at 90.5 percent for finding code vulnerabilities, at roughly half the cost per challenge.
- ExploitGym: frontier-level performance on validating which vulnerabilities are actually exploitable, using half the tokens Carbon needs.
- CWE-bench v2: 70.7 percent for patching vulnerabilities, a 3.4 percent uplift over Carbon.
The argument here is that as agentic coding scales, generating secure code from the start matters more than finding flaws later, and a model that writes safer code and finds and patches flaws faster is worth having even before it is generally available.
Safety and honest behaviour
Because Neon brings frontier-level capability to a faster, cheaper model, Google says it paired the release with real-time safeguards against cyber and biosecurity misuse comparable to Carbon's protections, while refusing far less legitimate work in those areas. It refused just 1.98 percent of tasks on Terminal-Bench 4.0 and none on Terminal-Bench-Science.
On behaviour, Google reports several internal measurements that are worth noting, while remembering that these are Google's own tests rather than independent ones:
- Cheating: Neon matched Carbon with 0.0 percent cheating on Terminal-Bench 4.0, and took shortcuts on just 2.45 percent of tasks on Google's coding cheating evaluation, the lowest rate of any Gemini 4 model.
- Misaligned actions: a 0.09 percent misaligned action rate on Google's agentic safety suite, and a 99.2 percent pass rate on its agent honesty evaluation.
- Overreach and under-service: compared with Carbon, Neon reduced under-service by 43 to 56 percent and overreach by 15 to 35 percent across different effort levels, meaning it is less likely to do too little or to go beyond what it was asked to do.
- Prompt injection: on Gray Swan's indirect prompt injection benchmark, where malicious instructions are planted in web pages, emails, documents or code the model accesses, Neon resisted every attack at its highest thinking level, matching the 0.0 percent attack success rate reached only by other Gemini models.
These figures are self-reported, and independent testing will matter. But they describe a model that Google has tuned specifically for the failure modes that show up in long agent runs: cutting corners, fabricating results and stepping outside its instructions.
Who should use it, and who should not
Neon is aimed at a fairly specific user, and Google is unusually clear about who that is.
- Developers and engineering teams running coding agents in production, especially multi-file work across repositories, will get the most from it. The number one ranking on the Coding Agent Index and the 65.2 percent on Terminal-Bench 4.1 are the figures that matter here.
- Enterprises doing knowledge work across spreadsheets, PDFs, charts and video, and teams that need an agent to work through long, multi-step jobs, will benefit from the 1 million-token context window and the lower cost per task.
- Security teams should watch the Fairwind rollout of Neon for cyber, which is aimed at finding, validating and patching vulnerabilities.
- Everyday users on the free Gemini plan will not see Neon at all. It is a paid-tier model, available to Google AI Pro and Ultra subscribers in the Gemini app and Antigravity.
- Anyone with the hardest single problem should still look at Carbon, which remains Google's model for the most demanding long-horizon work, where Neon's speed and cost advantages matter less than maximum depth.
Bottom line
Gemini 4 Neon is not Google's smartest model, and Google does not claim it is. It is the company's attempt to answer a different question: what does a frontier model look like when it has to do serious work, all day, at a price a business can afford? The answer, on Google's numbers, is a model that ranks third overall on the Intelligence Index, first among coding agents, and costs roughly a third less per task than Carbon while doing it. Whether those numbers hold up in independent use is the test that matters now, and it is one that developers and enterprises will run themselves in the coming weeks. For now, Neon is live for paid users, and it is the clearest sign yet that Google's AI strategy is no longer about one flagship model, but about a family of models tuned for different jobs, at different prices.
FAQs
What is Google Gemini 4 Neon?
Gemini 4 Neon is Google's AI model launched on November 18, 2026, built for complex, high-frequency work such as coding, analysis and science, with frontier-level reasoning at a lower cost than the flagship Carbon model.
How much does Gemini 4 Neon cost?
It costs $3 per million input tokens, $0.15 per million cached input tokens, and $15 per million output tokens, with a 1 million-token context window. Google says it costs 32 percent less per task than Carbon at its default thinking level.
How good is Gemini 4 Neon on benchmarks?
It scores 61 on the Artificial Analysis Intelligence Index, ranking third overall, ranks second on the Vals Index, and ranks first overall on the Artificial Analysis Coding Agent Index at 58.6.
Where can I use Gemini 4 Neon?
It is available through the Gemini API paid tier in Google AI Studio, Gemini for business, Google Antigravity and the Gemini app for paid Google AI Pro and Ultra subscribers. It is not available on the free Gemini plan.
What is the difference between Neon, Carbon and Hydrogen?
Carbon is for the hardest, longest projects, Hydrogen is for fast everyday tasks, and Neon sits between them, built for complex work done at high frequency, such as production coding and enterprise analysis.
What is Gemini 4 Neon for cyber?
It is a variant tuned for defensive security work, including finding, validating and patching code vulnerabilities. It is rolling out to trusted defenders through Google's Fairwind Program.
Was this article helpful?

Written By
TechorCoffeeContributing Writer
TechorCoffee is a technology reporter and software creator writing reviews and tutorials for TechOrCoffee.
Comments (0)
Leave a Comment
Read Next

Tue Oct 13 2026
Gemini 4 Argon: Everything You Need to Know
Gemini 4 Argon is Google’s new flagship AI model. See its benchmarks vs GPT-6 Astra and Claude, pricing, 1M-token output, safety features and how to get access.

Tue Oct 13 2026
Google Gemini 4 Argon: Features, Benchmarks, Pricing & Guide
Discover Google DeepMind's Gemini 4 Argon. Explore its 1M token output limit, #1 Vals Index ranking, API pricing, coding capabilities, and enterprise use cases.

Fri Oct 30 2026
Qwen 4 Is Not Out Yet: What Is Actually Confirmed
Qwen 4 in 2026: what Alibaba confirmed at Apsara, what the Flash-Next preview shows, what Kalmia rumours claim, and what is still unknown before launch.