Magai

Gemini 4 Neon: Price, Specs and Who Should Switch

Gemini 4 Neon is out and now in Magai. Specs, $3 and $15 pricing, benchmarks, where it loses, and who should switch.

By Dustin W. StoutPublished 17 min read
Three steel calibration weights in graduated sizes on a pale workbench, the middle one catching a blue reflection, standing for Gemini 4 Neon in the middle of the Gemini 4 family

Google just shipped the Gemini 4 model you will probably reach for most, and it launched without the one comparison that would prove it.

That model is Gemini 4 Neon, announced on November 18, 2026. It costs $3 per million input tokens and $15 per million output tokens.

Neon is the middle of the Gemini 4 family. Google adapted it from Gemini 4 Carbon, gave it a 1 million token context window and four thinking levels, and built it for complex coding, spreadsheet and document work, and long agent runs. Google pitches it as deeper than Gemini 4 Hydrogen and cheaper to run than Carbon.

Every number behind that pitch is Google's own.

Some of them are strong. One row is missing, and it's the row you need most.

So here is my verdict. Test Gemini 4 Neon this week on your own work, and don't move a single job off Carbon until your own numbers say it's safe.

The good news is that you can start today. Neon is now in Magai, and every Magai user can open it right away, beside the Claude and GPT models you already use.

What Gemini 4 Neon is, in plain terms

Gemini 4 is the generation. Argon, Carbon, Hydrogen and Neon are the models inside it, each named for an element.

Neon's model card says it is "adapted from Gemini 4 Carbon," which means the same architecture and training data as Carbon, tuned for a different job. Google's one line summary of that job: "While Carbon tackles your hardest, long-horizon projects and Hydrogen handles fast everyday execution, Neon balances depth and efficiency."

Think of three tool bags in graduated sizes. Hydrogen is the small bag by the door that handles most of the day. Carbon is the big rolling case you wheel out for the hard job. Neon is the medium bag, and the open question is how much of the big case it carries.

Here is what Google publishes about the model, in one place, plus the one row that matters most if you work in Magai:

Field What Google lists
API model ID gemini-4-neon
Announced November 18, 2026
Built from Gemini 4 Carbon
Context window 1 million tokens, shared across input, thinking and output
Output Up to 64K tokens per standard reply, up to 1 million with background long decode
Inputs Text, images, audio and video
Thinking levels Four, with high as the default and xhigh as the top level
Knowledge cutoff March 2026
Price $3 per million input tokens, $15 per million output tokens, $0.15 per million cached input tokens
Where it runs Gemini API paid tier in Google AI Studio, Gemini for business, Google Antigravity for paid Google AI Pro and Ultra subscribers, the Gemini app for Pro and Ultra subscribers
In Magai Available to every Magai user, from the model picker

Two details in that table deserve more attention than they get.

First, the 1 million token window is shared. According to the Gemini API docs, input, thinking and output all draw from the same 1,048,576 tokens. Paste in a 900,000 token document and you have about 148,000 tokens left for the thinking and the answer combined.

Second, output works differently than it did on older Gemini models. A standard reply tops out at 64K tokens, and anything longer goes through background long decode. If you plan to feed it a huge file and ask for a huge rewrite, plan for that split before you start.

The price looks steep until you count tokens

Start with the sticker prices side by side. All of these are per million tokens.

Model Input Output Note
Gemini 4 Hydrogen $1.50 $7.50 Released November 12
Gemini 4 Neon $3.00 $15.00 Cached input is $0.15
Gemini 4 Carbon $2.00, then $4.00 $10.00, then $20.00 Introductory rate ends November 30, new rate starts December 1
Gemini 4 Argon $2.00, then $4.00 $10.00, then $20.00 Same dates as Carbon

Read that table in two parts.

Through November 30, Neon costs 50% more per token than Carbon. From December 1, Carbon moves to $4 and $20, and Neon becomes 25% cheaper per token. Our Gemini pricing breakdown tracks every date where Google's numbers change.

Against Hydrogen, Neon is exactly double. That's the price of stepping up a tier, and it's only worth paying when Hydrogen stalls.

A small brass balance scale on a wooden desk, one pan holding coins and the other holding stacked blank receipts

Per token is the wrong unit anyway. You pay per task, and a model that thinks less or retries less can cost less even at a higher rate.

Here is the same job priced on each model, using list prices. The job reads 200,000 tokens and writes 30,000, thinking included. This is arithmetic on the published rates with identical token counts, not a measurement of any model.

Model Input cost Output cost Total
Hydrogen $0.30 $0.23 $0.53
Neon $0.60 $0.45 $1.05
Carbon through November 30 $0.40 $0.30 $0.70
Carbon from December 1 $0.80 $0.60 $1.40

Now the part the table hides. Models don't use identical token counts, and Neon is sold on using fewer.

Google's announcement says that at its default high thinking level, Neon costs 32% less per task than Carbon. It does not say which Carbon price that math uses, the introductory $2 and $10 or the $4 and $20 that starts December 1. That one detail changes whether the claim helps you in November.

Run the break even yourself. Through November 30, Neon has to use less than two thirds of the tokens Carbon uses to cost the same. From December 1, it wins on price unless it needs more than a third more tokens.

Against Hydrogen, the rule is simpler. If Neon clears in one try what Hydrogen needs two tries for, you break even. If Hydrogen clears it first time, you just paid double.

Thinking level is the other dial on your bill. Google gives Neon four levels and says each step up buys deeper multi-step reasoning for the hardest engineering, analytical and scientific work, with token spend scaling alongside it. High is the default, and xhigh is the top.

Google's cost claims attach to specific levels. At high, it says Neon costs roughly 50% to 80% less per task than the top competing frontier models. At xhigh, where it posts its 61 on the Artificial Analysis index, it says Neon still costs more than 36% less per task than competing models at their highest settings.

So test at high first. If a task fails there, rerun only that task at xhigh before you decide the model can't do it. The failure might be the setting, and you want to know what the fix costs.

Google's strongest claims are in knowledge work and agents

Google's announcement leads with rankings from outside evaluators, which is the right instinct. A score from a third party carries more weight than a score from the lab that trained the model. Here is the set it reports, with Google's own descriptions of where each lands.

Area Benchmark Neon's score Where Google says it ranks
Overall capability Artificial Analysis Intelligence Index at xhigh 61 #3 overall
Enterprise work Vals Index 70.9% #2
Agent sessions Arena Agent Arena Index 1336 #2 across 4,200+ real sessions
Workflow automation Zapier AutomationBench 56.3% #1
Business simulation Vending-Bench 2 $16,843 #1
Coding agents Artificial Analysis Coding Agent Index 58.6 #1
Science Terminal-Bench Science 70.5% Best in class
Math RiemannBench 84.0% Best in class
Hard reasoning HLE-Diamond 71.3% Best in class
Security SEC-bench Pro 90.5% State of the art

Steel calipers measuring a plain wooden block on a workbench, beside a ruler and pencil, a picture of checking Gemini 4 Neon's benchmark claims yourself

The pattern in that list matters more than any single row. The wins cluster in work where a model has to read a pile of material, hold a plan, and finish a long job: spreadsheets, legal and tax research, automation, agent sessions that run for hours.

Google also reports first place on its Legal Research Bench and Tax Agent Bench entries without publishing a score, and says Neon ranks first on most of Surge AI's finance, accounting, spreadsheet modeling and constraint following benchmarks. Those are the claims to test with your own documents, because a lawyer's contract and a benchmark's contract are different animals.

Vending-Bench 2 deserves a sentence of its own. It's a long running simulation of operating a business, and the $16,843 is what Neon ended with. It rewards a model that keeps its plan straight across many decisions, which is the exact skill you want in an agent that runs unattended.

Steerability is the claim I'd check first. Google says Neon ranks second on steerability, meaning how cleanly it adapts when you correct it halfway through a task. Anyone who has watched a model cheerfully ignore a correction knows how much that's worth.

The code result that tripled was the harness, not the model

Google reports that Neon ranks first on the Artificial Analysis Coding Agent Index, scores 65.2% on Terminal-Bench 4.1, and reaches 46.2% on FrontierCode 2.0 at its default high setting.

The SWE-sweep numbers are the ones to read slowly. Neon scores 12.8% on the mini-SWE harness and 39.0% inside Google Antigravity. Google says the score more than triples and credits the Antigravity harness for the jump.

Read that with the model held constant. Same weights, same benchmark, three times the score, and the only thing that changed is the scaffolding around the model: the tools it can call, how it reads files, how it handles long sessions.

That's a useful lesson for anyone shopping on benchmark tables. Where you run a model can matter as much as which model you pick.

It also says what to test. If your coding work happens in a chat window with pasted files, the mini-SWE number is closer to your experience than the Antigravity one. If your work happens inside an agent harness with tools, the higher number applies, and you should test it there.

Google also describes Neon acting as an orchestrator in Antigravity, directing parallel Hydrogen subagents. The idea is a medium bag that plans the job and hands the small tasks to the small bag. It's a sensible design, and it's the use case where Neon's price makes the most sense, since the expensive model only does the thinking.

Google's own table shows Neon losing in three places

The model card's results table puts Neon beside the newest models from Anthropic and OpenAI, and Neon doesn't win every row.

Neon trails Claude Opus on Terminal-Bench 4.1, 65.2% against 65.8%. It trails Claude Opus on FrontierCode 2.0, 46.2% against 47.3%. And it trails GPT Sol on Chartography, which tests reading dense professional charts, 73.5% against 75.1%.

Those gaps are about a point each, small enough that one unlucky run could swing them. Google deserves credit for printing them instead of dropping the rows.

Still, they say something. Neon is a strong generalist, and it isn't a clean sweep for terminal driven code or for chart reading. If your week is mostly command line work in a repo, the Claude flagship earns a place in your test. If your week is reading charts out of board decks, GPT does.

Here is a quick way to weigh the losses. Look back at your last ten tasks and sort them into command line or repo work, chart and image reading, and everything else. If seven of the ten are terminal work, a one point deficit on Terminal-Bench matters more to you than the knowledge work wins in the table above. If most of them are documents and spreadsheets, it matters much less.

A benchmark table is an average over other people's work. Your ten tasks are the only sample that counts.

The honest verdict on rankings: Neon is in the top tier on nearly everything Google reports, and first on the work it was built for. That's a strong result. It isn't the same as best at your job.

Google never puts Neon's scores beside Carbon's

Here is the missing row.

In the text of the announcement, Google compares Neon with Carbon on cost per task, on a handful of security tasks, and on agent behavior. It does not put Neon's general capability scores in a column next to Carbon's. The model card runs its safety comparison against Hydrogen instead.

The security comparisons are specific. On CWE-bench v2, Neon reaches 70.7%, a 3.4% uplift over Carbon. On ExploitGym it matches frontier level performance using half the tokens of Carbon. On AutoBaxBench, Neon produces more secure code than Carbon at every difficulty, and Google reports 76% on the hard tier.

On agent behavior, Google says Neon cuts under-service by 43% to 56% and overreach by 15% to 35% compared with Carbon, depending on the effort level.

Those are real comparisons. They're also all in the areas where Google chose to publish one.

So the question that decides whether you move work from Carbon to Neon has no official answer yet: how much capability do you give up for the lower cost per task? Google's framing is that you give up little. Google is also the one selling both models.

You might object that a model adapted from Carbon, and priced below it after December 1, must be a slightly smaller Carbon, and that's all you need to know. Fair enough, as far as price goes. It doesn't tell you where the smaller version breaks, and the break point is the whole decision.

If you pay for Carbon today, you're the person with the most to gain from running this comparison, and the most to lose from skipping it.

Pick between Hydrogen, Neon and Carbon by where your work stalls

Google's own guidance is a ladder, and it's a good one. Start low, move up when the cheaper model fails you, and stop moving when it stops failing.

Three canvas tool bags in small, medium and large sizes on a concrete floor, the medium bag open with a blue strap

Our posts on Gemini 4 Hydrogen and Gemini 4 Argon cover the other two ends of the family in detail. For the decision in front of you now, match your situation to a row:

If you are Do this
Wanting to test Neon on your own work today Open it in Magai, from the model picker or with /Gemini, and run your tasks against Claude and GPT in the same thread.
Running everyday coding and parallel worker agents on Hydrogen without trouble Stay. Google positions Hydrogen for fast everyday execution, and Neon doubles the per token price.
Hitting Hydrogen's wall, with retries and wrong answers on spreadsheets or long documents Test Neon first. If it clears the task in one try, the price difference disappears.
Paying Carbon's introductory rate on long agent jobs Wait until December 1 to compare costs, and run your own tasks before you move anything.
Orchestrating many subagents Test the pairing Google describes, Neon planning and Hydrogen executing.
Doing legal, tax, finance or spreadsheet work Neon's claimed first place rankings sit in these areas, so test with your own documents.
Using Gemini casually in the app Nothing to do. Neon appears in the model dropdown for Pro and Ultra subscribers, and Hydrogen handles most of what you ask.
On a security team Google is rolling out a cyber version of Neon to trusted defenders through its Fairwind Program, so general access isn't the path.

The tool bag picture helps here. You don't buy the medium bag so you can leave the small one at home. You buy it for the jobs where the small one runs out of room.

Take the orchestration row and carry it to a number, since Google's pitch for Neon leans on it. The token counts below are assumptions I picked to make the arithmetic visible, not measurements. The job has one planning pass of 40,000 input tokens and 8,000 output tokens, then five execution runs of 30,000 input and 6,000 output each.

Setup Planning pass Five execution runs Total
Everything on Hydrogen $0.12 $0.45 $0.57
Neon plans, Hydrogen executes $0.24 $0.45 $0.69
Everything on Neon $0.24 $0.90 $1.14

Putting Neon on the planning pass alone adds $0.12 to the job, about 21% over an all Hydrogen run. Putting it on every run doubles the bill.

That's the case for the pairing. You buy the better judgment where one decision steers five runs, and you skip it where the work is routine.

The failure mode is a cheap executor that makes a mistake the planner never sees. Fix it by making the last step a verification pass that reads the actual outputs, and give that pass to Neon too.

The safety numbers are Google's, so check them with a trap

Google puts real weight on reliability in this launch, and it's the part I'd read twice.

The announcement says Neon records Google's lowest fabrication rate yet in internal testing in Antigravity. On Google's coding cheating evaluation, it took shortcuts on 2.45% of tasks, the lowest of any Gemini 4 model. It reports a 0.09% misaligned action rate on its agentic safety suite and a 99.2% pass rate on its agent honesty evaluation. On Terminal-Bench 4.0 it refused 1.98% of tasks.

On prompt injection, Google says Neon resisted every attack in Gray Swan's benchmark at its highest thinking level, with real-time safeguards switched on.

All of that is internal or vendor run. The model card says the full Safety and Security report is "coming soon," and it also states plainly that Neon "may exhibit some general limitations such as hallucinations." Its knowledge cutoff is March 2026.

So don't read 99.2% as a promise about your files. Read it as a reason to run one specific test. Give the model a task built on a false premise, ask it what the early termination clause says in a contract that has none, and see whether it invents one.

A model that says "there is no such clause" has earned some trust. A model that quotes a clause that doesn't exist has told you everything you needed to know.

Run five of your own tasks in Magai before you move anything

Gemini 4 Neon is in Magai, and every Magai user can open it today. Pick it from the model picker, or start a message with /Gemini and choose Neon.

That matters for this test in one way. You can send Neon, a Claude model and a GPT model the same task in one Magai thread, with your files and earlier messages intact, and score them side by side. The thread keeps its memory when you switch models, so no model gets a cleaner run than the others.

Five blank index cards taped in a row on a plaster wall above a tidy desk, with a blue tab on the middle card

The five tasks below come straight from Google's own claims. Here is the setup, in order:

  1. Open a new thread and attach your files once: a few files of real code with a bug in them, a spreadsheet export and a PDF from the same project, a very long document, and a screenshot of a dense chart from your own reports.
  2. Start the first message with /Gemini, choose Neon, and type the task from the table below.
  3. Switch the same thread to /Claude, then /GPT, and send the same task. The thread keeps what came before, so each model sees the same files and the same corrections.
  4. Score each answer from 0 to 2 against the pass description, and keep a running total per model.
  5. When the fifth task is done, compare the totals. Neon has earned a job only where it ties or beats the model you use for that job today.

The five tasks, with the Google claim each one tests:

Task Google's claim A pass looks like
Fix a bug across several files Targeted, surgical edits across repositories The smallest diff that fixes the bug, with no other files touched
Build a summary from a spreadsheet and a PDF Turns inputs into structured deliverables Every figure traced to a file, and any disagreement between the files named
Answer questions from the middle of a very long document A 1 million token window Exact quotes, correct answers, nothing invented
Read a dense chart 73.5% on Chartography Five correct values, and a plain list of which ones it estimated
Take a correction halfway, then handle a false premise Second on steerability, lowest fabrication yet It changes only what you corrected, and it says when the thing you asked about isn't there

The prompts, ready to paste. The first two use your own files:

Here are three files from my project. The bug: [describe the symptom].
Find the root cause and change as few lines as possible. Show each
edit as a diff. Do not rewrite any file.
Using only the attached spreadsheet and PDF, write a one page summary
with three findings. Give each finding's number and the file it came
from. If the two files disagree, say so instead of picking one.

For the long document, ask three questions whose answers sit in the middle pages, and add this line:

Quote the exact sentence that supports each answer. If you can't find
one, say you can't find one.

For the chart, attach the screenshot and send:

Read off five values from this chart and describe the trend. Then tell
me which of the five values you estimated instead of read directly.

For the last task, send two messages. The first gives the model a real piece of work, a plan with a budget, say. The second reads:

Correction: the budget dropped by $12,000. Redo only the budget
section and leave everything else exactly as it was.

Then, in a fresh message, ask it about a clause that your contract doesn't contain.

Expect to see differences in two places first. Some models rewrite everything after a correction, and some confirm a false premise politely. Both failures cost you real time in real work, and neither shows up on a leaderboard.

Then apply one rule when you read the totals. If a cheaper model ties a pricier one on four of the five tasks, the pricier one has to earn its place on the fifth. That rule applies to Neon and to every other model in the picker. Our posts on the best AI chatbot and Claude vs Gemini use the same task first approach for the other choices you're weighing.

If you're not on Magai yet, plans start at $20 a month, listed on the pricing page, and every plan carries a 30 day, 100% money back guarantee.

Keep the medium bag in the truck. Open Magai, run the five tasks, and let Neon audition for the job where the small bag runs out of room.

The big case can wait until the math changes on December 1.

More in Artificial Intelligence

  • A balance scale on a dark desk with a heavy steel weight on one side and a small light weight on the other, level, with a single blue object beside it

    Artificial Intelligence

    GPT-6.2 Astra and Sol: Who Should Switch

    GPT-6.2 launched November 17 as two models, Astra and Sol. Specs, API prices, benchmarks, where it slips, and who should switch.

  • A small blue glass flask beside a heavy steel anvil on a dark workshop bench, lit by one soft light

    Artificial Intelligence

    Gemini 4 Hydrogen: What It Is and Who Should Use It

    Gemini 4 Hydrogen is Google's fast, cheap Gemini 4 model, out November 12. Specs, price, benchmarks, where it falls short, and who should switch.

  • A single worn engine part resting on a white shop towel on a dark mechanic's workbench under one lamp

    Artificial Intelligence

    Perplexity vs ChatGPT: Pick by How They Fail

    Perplexity vs ChatGPT: one cites sources you can click, the other sounds sure even when wrong. How each fails, what each costs, and a 20 minute test.