Key Takeaways
- Neon Joins Gemini: Google released Gemini 4 Neon on November 18, 2026, placing the reasoning model between Hydrogen and Carbon. If you use the Gemini app, Neon replaces Argon under Deeper analysis for AI Pro and Ultra subscribers.
- Access and Costs: Gemini 4 Neon’s standard API costs you $3 per million input tokens and $15 per million output tokens, including thinking. There is no free API tier, and Google hasn’t announced free or AI Plus app access.
- Test Before Switching: Most benchmark scores are Google’s own, so test Gemini 4 Neon on familiar work before switching. If you develop apps, try Batch/Flex at $1.50 input and $7.50 output per million tokens and compare completed-task costs.
If you pay for Gemini, you now have another model to choose for work that needs more thought. Google released Gemini 4 Neon on November 18, 2026, placing it between Hydrogen and Carbon. Neon replaces Argon in the Gemini app and is rolling out to Google AI Pro and Ultra subscribers.
For developers, paid API access starts at $3 per million input tokens and $15 per million output tokens. The pitch is better reasoning for coding, document analysis and other demanding jobs, with less time and money spent completing each task.
Here’s the thing. Google hasn’t said whether free or AI Plus subscribers will get Neon in the app, and there is no free API tier. Most benchmark scores supporting the launch are Google’s own, too. Before you pay or change models, here are the details that matter.
| Detail | What you need to know |
|---|---|
| Release date | November 18, 2026 |
| Model ID | gemini-4-neon |
| Place in lineup | Between Hydrogen and Carbon |
| Replaces | Argon in the Gemini app picker |
| Gemini app access | Google AI Pro and AI Ultra |
| API price | $3 input / $15 output / $0.15 cached input per million tokens |
| Context window | About 1 million tokens |
| Thinking levels | low, medium, high by default, xhigh |
| Knowledge cutoff | March 2026 |
| Free tier | No free API tier. Google hasn’t stated app access for free users |
What Google Built Gemini 4 Neon to Do
Google built Neon for jobs where you need more reasoning than a quick reply usually requires. That includes debugging software, working through business documents and tackling scientific problems. Its position between Hydrogen and Carbon gives you another option for demanding tasks that you repeat regularly.
According to DataCamp’s account of the model card, Neon was adapted from Carbon. Google’s aim is to retain strong reasoning while reducing the time and expense needed to finish a job. That matters because a cheaper individual response is less useful if you need several attempts to get usable work.
Inputs and Thinking Levels
You can provide text, images, audio and video, with PDF analysis and Office or OpenDocument file support on Google Cloud. Neon returns text, so those input capabilities do not mean you can generate pictures or audio.
You also get four thinking levels, with high selected by default and xhigh available for deeper reasoning. To understand when you might choose Neon, start with the models around it.
Where Neon Sits in the Gemini 4 Family
Hydrogen handles quick tasks, Carbon targets the hardest projects, and Neon occupies the space between them. Argon is leaving the app picker as Neon takes its place. Here’s how those roles and standard API prices compare.
| Model | Intended role | Gemini app picker label | API input / output per million tokens |
|---|---|---|---|
| Hydrogen | Quick responses and routine execution | Everyday help | $1.50 / $7.50 |
| Neon | Demanding analysis, coding and knowledge work | Deeper analysis | $3 / $15 |
| Carbon | Difficult projects requiring sustained reasoning | Code & complex work | $2 / $10 through November 30, then $4 / $20 |
| Argon | Earlier model being replaced in the app | Being removed from the picker | $2 / $10 through November 30, then $4 / $20 |
You may also see Auto, which 9to5Google identifies as the paid-user label for Gemini 3.5 Flash-Lite. The same report says Gemini 3.1 Pro is being removed, so the picker changes extend beyond Argon.
That said, a label appearing in the lineup does not guarantee access on your plan. According to Android Authority, you need an Ultra subscription to choose Carbon in the Gemini app, whereas you can use Neon with either Pro or Ultra. That makes your subscription the next thing to check.
Which Gemini Plans Get Neon
Google names AI Pro and AI Ultra as the consumer subscriptions receiving Neon. App access is arriving gradually, so you may not see the new choice immediately. Your route depends on whether you use the Gemini app, development tools or a business account.
In the Gemini App
- AI Pro and AI Ultra. Google confirms Neon access for both subscriptions.
- Free and AI Plus. Google hasn’t said whether either group will receive Neon. Neither is included in the announced app rollout.
- Gradual availability. Your account may receive the picker change later than another subscriber’s account.
Once Neon reaches your account, you can select it in three steps.
- Open Gemini.
- Tap the model picker.
- Choose Deeper analysis.
In Antigravity and for Business
- Google Antigravity. Paid AI Pro and Ultra subscribers are receiving Neon through the desktop app, CLI and IDE extensions.
- Gemini for business. Google includes its business offering among the announced access routes.
- Google Cloud. The model page lists global, US and EU multi-region availability, with access for verified enterprise customers in selected regions.
Google’s Antigravity announcement does not provide numerical Neon usage limits. Before you build a regular workflow around it, check the access and limits shown for your account.
How Much Neon Costs Developers
Neon’s standard API rate is $3 per million input tokens and $15 per million output tokens. Cached input has a lower rate, and DataCamp lists Batch/Flex and Priority pricing alongside the standard option. Here’s what each tier costs before you compare Neon with Carbon.
| API tier | Input per million tokens | Output per million tokens | Cached input per million tokens |
|---|---|---|---|
| Standard | $3 | $15 | $0.15 |
| Batch/Flex | $1.50 | $7.50 | $0.075 |
| Priority | $5.40 | $27 | $0.27 |
Thinking tokens count toward your output bill. So a short visible answer can still cost more than you expect if Neon uses substantial reasoning before producing it. Compare the total cost of completed jobs, rather than judging expense from answer length alone.
Neon Versus Carbon and Hydrogen
The Carbon comparison needs a date attached. Through November 30, Carbon’s promotional $2 input and $10 output rates are below Neon’s standard rates. From December 1, Carbon rises to $4 and $20, making Neon’s corresponding rates 25% lower.
Hydrogen remains cheaper at $1.50 input and $7.50 output per million tokens. That puts Neon at twice Hydrogen’s standard token prices, although your final task cost also depends on how many tokens each model uses.
Caching and Lower-Cost Testing
There is no free Neon API tier. Implicit caching requires at least 4,096 prompt tokens, and explicit context caching is unsupported. Google Cloud separately lists cache writes at 1.25 times standard input pricing, so the discounted cached-input rate does not describe every cache-related charge.
If your jobs can wait, Batch/Flex gives you a lower-priced starting point. Test the same representative tasks before moving regular traffic, and include thinking tokens when you compare the bills.
What the Benchmarks Say
Google’s results suggest Neon can handle demanding coding and business tasks, but most launch scores are Google’s own. You should treat them as evidence to investigate, rather than a promise about your work. Start with the reported numbers, then look at the comparisons behind them.
| Benchmark | Google-reported Neon score |
|---|---|
| Artificial Analysis Intelligence Index | 61 at xhigh |
| AA Coding Agent Index | 58.6 |
| Vals Index | 70.9% |
| AutomationBench | 56.3% |
| Terminal-Bench 4.1 | 65.2% |
| FrontierCode 2.0 | 46.2% at high |
| HLE-Diamond | 71.3% |
Comparisons and Independent Results
Neon does not win every comparison. Google’s tables put Claude Opus 5.5 ahead on Terminal-Bench 4.1 at 65.8% and FrontierCode 2.0 at 47.3%. GPT-6.2 Sol leads Neon on Chartography, scoring 75.1% against 73.5%.
There is independent context, too. Lindy’s summary of Artificial Analysis testing puts Neon one point above Carbon, with an Index task costing about $6.44 against Carbon’s $7.76. That is roughly 17% less, a saving also noted by the Free Press Journal.
That said, a small index lead cannot tell you which model will handle your spreadsheet or codebase better. Use these results to choose what to test, then judge the finished work yourself.
What Changes for Everyday Gemini Users
If you already pay for Pro or Ultra, the useful change is another option for heavier work. Google says Neon improves writing, document reading and tasks that require several steps. Here’s where you might try it first.
- Writing and organization. Ask Neon to revise a report or turn scattered notes into a structured response.
- File analysis. Use spreadsheets, PDFs or photos when your question depends on details inside a document.
- Current research. Ask for web research when you need information newer than the model’s built-in knowledge.
- Connected documents. Neon can draw details from Gmail and Docs for work involving your connected information.
- Tasks in Gemini. Google gives billing-statement analysis and receipt cross-checking as examples of work Neon can run.
Keep Everyday help for quick requests and reserve Deeper analysis for demanding jobs. Check your account’s usage limits before making Neon your regular choice for everything.
What Developers Should Know Before Switching
Changing the model ID is only part of the move. Neon has controls and limits that can affect an existing integration, especially if you reuse parameters from another model. Check these points before sending regular traffic.
- Model and thinking. Use `gemini-4-neon` with `thinking_level` set to `low`, `medium`, `high` or `xhigh`. High is the default, and `minimal` is rejected.
- Sampling parameters. Valid `temperature`, `top_p` and `top_k` values are ignored. Out-of-range values can produce a 400 error.
- Output and timeouts. Synchronous output stops at 65,536 tokens. Lindy recommends client timeouts of at least 15 minutes.
- Caching and Cloud support. Explicit caching is unavailable. Google Cloud also lists tuning, RAG Engine and Gemini Live API as unsupported.
- Antigravity orchestration. Neon can plan the work while Hydrogen performs routine edits and tests, returning difficult steps to Neon.
That planner-and-executor arrangement gives you another way to test costs. Compare it with using Neon alone on representative jobs before deciding which workflow to keep.
Safety and the Neon for Cyber Version
Google reports a 0.09% rate of actions it classified as misaligned and a 99.2% pass rate on agent honesty testing. Those are test results, so you should still review consequential actions and completed work.
At launch, Google said trusted defenders would receive the cyber-specific Neon variant in the days following the announcement. Access is through the Fairwind Program, with interested organizations directed to their Google Cloud account team.
There is an unresolved qualification. According to Lindy, Google carried over Argon’s safety determinations, and the complete Safety and Security Report was still pending. Keep that distinction in mind when assessing Google’s safety claims.
Should You Pay for Gemini 4 Neon
If you use Gemini for free, wait for a clear access announcement. Google hasn’t said whether free or AI Plus users will receive Neon, so neither group should assume access.
If you already have AI Pro, you have the clearest reason to try Neon on demanding documents, writing or coding. Developers can start with Batch/Flex and compare completed-task costs before committing more traffic.
Give Neon one job you know well, then check its accuracy, completion time and cost. You’ll have a useful basis for deciding where it belongs in your work as access rolls out.
FAQs
You cannot use Neon through a free API tier. Google hasn’t said whether free Gemini app users will receive access.
You need Google AI Pro or AI Ultra under the announced rollout. Choose Deeper analysis in the model picker once it appears for your account.
You cannot assume Neon is better for every task, although independent Artificial Analysis results put it slightly ahead overall. Carbon targets difficult, sustained projects, so compare both on work you actually need completed.
You pay standard rates of $3 per million input tokens and $15 per million output tokens, including thinking tokens. Cached input costs $0.15 per million tokens, while Batch/Flex offers lower rates.
Yes, Neon takes Argon’s place in the Gemini app under Deeper analysis. The app rollout is gradual, so your picker may not change immediately.
You get a 1,048,576-token context window shared by input, thinking and output. That does not mean you can receive a response that large, because synchronous output stops at 65,536 tokens.

