Article · 9 min read

Gemini 4 Neon Is Google's Best Model

Google released four models in seven weeks, and the "middle" one comes out on top. Neon offers more than the flagship Carbon, a task run on it costs significantly less than on Opus, and people rate its writing among the best. In the first independent coding test, though, it trails Opus by seven points. Who is it worth it for, and how does the rest of the Gemini 4 family stack up?

Google released Gemini 4 Neon on November 18, which by its own description is a model for complex programming, analysis, and enterprise agentic work. I’m writing about it because, after Argon, Carbon, and Hydrogen, the Gemini 4 family is finally complete and can be measured against OpenAI, Anthropic, and the Chinese models. And because Google doesn’t come out of that comparison the way it would like.

What Neon is

A Carbon derivative at a lower price. According to the model card, it was created by modifying Carbon. A million input tokens cost $3, output tokens $15, and cached input $0.15. Compared with Carbon after the launch discount ends (from December 1), that is a quarter less; compared with Sonnet 5.5 and GPT-6.2 Sol, it is half again as much.

Four thinking levels. Low, medium, high (the default), and xhigh, with a context window of one million tokens.

Where to find it. In the paid Gemini API and AI Studio, in Antigravity (Google’s agent-based programming environment) for Pro and Ultra subscribers, and in the Gemini app. There, according to 9to5Google, it replaced both Argon and the old Gemini 3.1 Pro. The lineup in the app is now Hydrogen for everyday use, Neon for deeper analysis, and Carbon for code and the hardest work.

Independent measurement

ModelAA indexCost per task (USD)Price per 1M tokens input/output (USD)
Top models
Fable 5.2 (max)6526.548/40
Fable 5.2 (high)629.088/40
GPT-6.2 Astra ★624.008/40
Gemini 4 Carbon (high)605.48 (from Dec. 1: 7.76)2/10 (from Dec. 1: 4/20)
Opus 5.55910.044/20
Work models
Gemini 4 Neon (xhigh)616.443/15
Gemini 4 Neon (high)603.743/15
GPT-6.2 Sol ★571.012/10
Sonnet 5.55510.022/10
Kimi K3.1 Pro463.682/10
Mid models
Gemini 4 Hydrogen (xhigh)543.161.50/7.50
Gemini 4 Hydrogen (high)511.871.50/7.50
GPT-6.1 Terra ★510.961/5
Cheap models
Haiku 5.5 ★360.410.10/0.50
DeepSeek V4.1 Flash33not reported0.30/1.20
Gemini 3.8 Flash31not reported0.75/3.75
MiMo-V2.6-Pro30not reported0.43/0.87
GPT-6 Luna280.150.10/0.50

How to read the table. The AA index is the score in version 5.0.1 of the index from Artificial Analysis (an independent firm that measures models on its own set of tasks), higher is better. Cost per task is the average cost of one task in the index. The star ★ marks the best-value model in each tier, meaning the most points per dollar without a significant loss of performance. Models run at their highest thinking level unless the name says otherwise. For Carbon, the figure in parentheses is the price after the launch discount ends on December 1.

Neon works on one task for 21 minutes at xhigh and 15 minutes at high; Carbon takes 37 minutes.

Third in the world, first at Google. At the highest level, Neon scores 61 points, with only Fable 5.2 and GPT-6.2 Astra ahead of it. It edged past the flagship Carbon by a point.

High is the right level. It gives 60 points, the same as Carbon, for half of Carbon’s post-discount price and 2.5 times faster. A task comes to roughly 85 Czech crowns (about $3.70). At xhigh you get one more point for 1.7 times the price and six more minutes of work, which doesn’t pay off.

It loses to OpenAI in this index. Astra gives more points for almost the same price as Neon at high, and finishes faster. Sol gives 57 points for a dollar. Anyone looking for the cheapest finished work will still find it at OpenAI.

On the Vals Index it’s the opposite. In the Vals AI test of work in finance, law, taxes, and programming, Neon is second with 70.9 percent, behind Fable 5.2 (71.3) and just ahead of Sonnet 5.5, Carbon, and Opus. The top five models are tied within the statistical margin of error, but Neon is the cheapest of them: the test costs $7.40 on it versus $24.77 on Opus. Astra is only eighth here. On the finance and law index, Neon is first.

How does Gemini 4 handle Czech text?

Google’s models are traditionally said to be very good with Czech text (well, I’m the one saying it, it really does seem that way to me). So I have a game for you. You will always get two texts devoted to a single mobile network feature. I asked all the models in the Gemini 4 family to generate a short article about the evolution of this feature, 900–1,200 characters long, based exclusively on source materials supplied in advance. The point of the test was to measure how well the models stick to the sources and how readable the resulting text is. Plus, as it turned out, to evaluate how many things they made up beyond the sources, and which ones they invented entirely. If that doesn’t interest you, skip straight to the results.

I won’t keep you in suspense: on technical texts, Neon (and the other Gemini models and Sonnet) do very well; the one that came up short was Luna.

Programming: a draw in the tables, Opus in long-running work

Coding agent index. Artificial Analysis measured Neon in Antigravity at 59 points, and Fable 5.2 and Opus 5.5 in Claude Code at 58 each. That is the same level, but a task on Neon costs $6.86 versus $12.61 for Opus and $15.47 for Fable. In its announcement, Google turns this into first place. With a one-point difference and different tools, I wouldn’t call it that.

Google’s original table reveals more than it probably wanted to. On Terminal-Bench 4.1 and FrontierCode 2.0, Opus 5.5 scores slightly higher than Neon. Neon has a clear lead only in the SWE-sweep test (39 versus 10 percent), but Google measured that in its own Antigravity. In the standard environment for all models, Neon gets 12.8 percent. And one more thing: both Fable 5.2 and Astra are missing from the table, that is, both models that rank higher in the independent index. Well, these things happen…

The first independent test of long-running work. The company Bito, which develops a code review tool, ran Neon in Claude Code on the same long tasks as the other models:

ModelOverall scoreCost per sessionTime vs. Opus
Claude Opus 5.586.9%$18.38same
Gemini 4 Neon79.8%$7.3739% faster
Claude Sonnet 5.572.3%$9.7628% faster
Claude Haiku 5.570.9%$4.7045% faster

On one-off questions about code, Neon was actually slightly better than Opus. It lost out in a twenty-turn session and wherever it confidently made up how a library behaves. In one task it averaged one misleading claim per answer, while Opus had one per five answers. Opus, meanwhile, ran only at its default medium level; at a higher one, its lead would probably grow even more.

What follows from this. Compared with Sonnet, Neon is both better and cheaper. For long, autonomous work where you don’t have time to check every claim, Opus remains the safer choice.

Where Google is good: writing and the office

People like Gemini 4’s texts. In the Arena, where people blindly pick the better of two answers, Gemini 4 models took three of the top four places in the creative writing category: Argon first, Carbon third, Neon fourth. Opus 5.5 is second, and the differences are within the statistical margin of error, so it’s more of a shared first place. For marketing, emails, and everyday texts, that is a strong argument for Google. That is English; I wrote about Czech above, and you can try it yourself.

Best at presentations. In Artificial Analysis’s evaluation of work outputs, Neon leads in presentation quality, overtaking the previous leader, Sonnet 5.5.

One caveat. The EQ-Bench leaderboard, where another model grades short stories against a scoring rubric, ranks Google far below OpenAI, Claude, and even the Chinese Kimi. The crowds in the Arena reward fluent, friendly text, and measure literary quality less well. And we are measuring Czech together, on one type of task, processing supplied texts…

Where to be careful

It talks a lot and is slow to get going. During the index measurement, Neon produced 283 million output tokens, nearly the most of any model measured. At xhigh, you wait over 50 seconds for the first word of the answer. For chat, where you’re waiting for a quick reply, xhigh is not suitable at all.

Antigravity is reporting problems. One user on the Google developer forum documented that since the evening of November 18, all Gemini 4 models in Antigravity think 3 to 4 times less and the xhigh level returns an error. The success rate of his tasks dropped from 169 out of 190 to 2 out of 24. Google hasn’t confirmed this yet, but if you’re trying Neon in Antigravity, keep an eye on your results.

What Google is planning next

Krypton and Helix. Two more names are circulating on X: Krypton, allegedly a Fable competitor planned for December, and Helix, a cheap Flash-Lite-class model with a leaked price of $0.30 and $2.50. Google hasn’t confirmed either. Helix would fill precisely the hole in the cheap tier.

Take leaks with a grain of salt. When Neon appeared in the code of the Gemini website on November 16, the media guessed it would be a cheap Flash-Lite. Two days later it was released as the second most powerful model in the family.

Bottom line: what to take away

If you use Sonnet 5.5 through the API, try Neon at the high level. According to the first independent test, it is both better and cheaper.

For long autonomous programming, stick with Opus 5.5. Neon is faster and cheaper, but it more often confidently makes things up.

For writing, presentations, and work with documents, Neon is a strong choice, especially if you work in the Google ecosystem.

GPT-6.2 Sol still gives you the cheapest finished work. For cheap tasks, take Haiku, Luna, or DeepSeek; Google has nothing here.

Drop Argon, and buy Carbon only where you can verify its advantage yourself.

In my view, Neon is the first Google model in a long time that I would recommend as the default for everyday demanding work. But the Gemini 4 family feels as if Google were assembling it on the fly: four models in seven weeks, two of them already without a clear purpose, the best one being the middle one, and not even users can make sense of the names. Anyone who remembers Flash and Pro knows this used to be clearer.

Sources

Google’s announcement and the Gemini 4 Neon model card from November 18, 2026, measurements by Artificial Analysis (index version 5.0.1, coding agent index, cost and time per task), Vals AI (Vals Index as of November 18), the Bito test, Arena (creative writing), EQ-Bench, 9to5Google, the Google AI developer forum, i10x, and TestingCatalog (leaks). Google’s tables with Terminal-Bench, FrontierCode, and SWE-sweep are measured by the vendor. Exchange rate: 1 USD ≈ 23 CZK. Data valid as of November 19, 2026. The text was written with the assistance of Claude Opus 5.5, that is, a competing model.

Patrick Zandl trains teams in Claude Code and agentic software development in Prague and writes about what actually works in real-world AI-assisted projects. Follow on LinkedIn

Read next

Get the next one by email

Every Tuesday: what changed in Claude Code, Codex and other coding agents, plus new articles. No spam.