Gemini 4 Neon: Google's top-tier model for $3
Gemini 4 Neon is Google's new model: 61 points on Artificial Analysis, a 1M context window and $3 per million input tokens. It is in the API now. Here is what it can do.
This article in other languages:Polski
In this article
On November 18, Google released Gemini 4 Neon, the fourth model in the Gemini 4 family. Koray Kavukcuoglu (Google DeepMind, Chief AI Architect at Google) announced it as a top-tier model for coding, office work and science that also gets tasks done quickly. The same day, Neon started rolling out to the Gemini API, AI Studio, the Gemini app and Google Antigravity.
What is new here from an AI point of view?
Until now, the Gemini 4 family had two poles. Carbon handles the hardest tasks, the ones that run for hours, and Hydrogen handles fast, everyday execution. Neon is meant to sit in the middle: according to the announcement, it combines depth with efficiency on the complex tasks we do often.
Except that, going by the numbers, Neon does not look like a “middle” model. On the Artificial Analysis Intelligence Index, it scores 61 points at the xhigh thinking level and ranks 3rd overall. Carbon had 59 points when it was announced in October, and that put it first. Three weeks later, a score two points higher is only good enough for the podium.
In the public leaderboards Google cites, Neon is 2nd on the Vals Index (70.9%) and 2nd on Agent Arena, with a score of 1336 after more than 4,200 agent sessions. It scores 65.2% on Terminal-Bench 4.1 (Hydrogen scores 57%) and 46.2% on FrontierCode 2.0 at the default high level.
There are four thinking levels, the same as in Hydrogen. The default is high, and xhigh is intended for the hardest engineering, analytical and scientific tasks.
How much does Gemini 4 Neon cost?
The rates on the Gemini API pricing page look like this (USD per million tokens):
| Model | Input | Output |
|---|---|---|
| Gemini 4 Hydrogen | 1.50 | 7.50 |
| Gemini 4 Neon | 3 | 15 |
| Gemini 4 Carbon (through November 30) | 2 | 10 |
| Gemini 4 Carbon (from December 1) | 4 | 20 |
Neon costs exactly twice as much as Hydrogen. Cached input tokens are $0.15 per million, which is 95% cheaper, and in Batch and Flex modes the whole rate drops by half. The context window is 1M tokens.
The table also shows something the announcement leaves out: until the end of November, Neon is more expensive than Carbon, which still has its promotional price. Only from December 1 does the lineup fall into the order the names suggest, with Neon coming out a quarter cheaper. Google says that at the high level a task done by Neon costs 32% less than the same task done by Carbon, but it does not say which Carbon price that comparison uses.
How do you try Neon?
- In the Gemini app, pick it from the model list if you have a Google AI Pro or Ultra subscription.
- In the Gemini API and AI Studio, it is available under the name
gemini-4-neon, on the paid tier only. Unlike Hydrogen, it has no free tier in the API. - In Google Antigravity, it requires a paid Google AI Pro or Ultra plan.
- In Gemini for business, it works through the API and in the model list.
In Antigravity, Neon can also work as the Orchestrator, splitting tasks between parallel subagents that run on Hydrogen. In the Orchestrator Google showed alongside Hydrogen, the planning was done by Carbon or Argon.
A separate variant, Gemini 4 Neon for cyber, will reach trusted companies and institutions in the Fairwind program in the coming days. Google previously offered Carbon through the same program.
What should you watch out for?
First, the harness behind the coding results. On SWE-sweep, Neon scores 12.8% in the simple mini-SWE harness and 39.0% in Google Antigravity. Same model, a gap of more than three times. So the Antigravity number describes Google’s product, not what you get by plugging Neon into your own AI agent through the API.
Second, “best in its class”. That is how Google describes the results on Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%), but, as with Hydrogen, it does not name the models it compared against.
Third, the safety numbers, because they come from Google’s own tests. Neon is said to have the lowest rate of fabricated claims measured so far in Antigravity’s internal tests, to score 99.2% on an agent honesty test and to allow 0.0% successful indirect prompt injection attacks on the Gray Swan benchmark at the highest thinking level. The details are in the model card.
Fourth, the pace. Argon came out on October 13, Carbon on October 28, Hydrogen on November 12 and Neon on November 18. Four models in a little over a month means plenty of choice, but also a fair amount of confusion: the most repeated complaint in the Hacker News comments is that the element names tell you nothing about which model is for what :)
FAQ
Is Gemini 4 Neon free?
No. In the Gemini app and in Antigravity it requires a Google AI Pro or Ultra subscription, and in the API it is available on the paid tier only: $3 per million input tokens and $15 per million output tokens.
How is Gemini 4 Neon different from Carbon and Hydrogen?
Carbon is for the longest and hardest tasks, Hydrogen is for fast execution. Neon is meant to sit between them, although it scores 61 points on the Artificial Analysis Intelligence Index, two more than Carbon had at its announcement.
What context window does Gemini 4 Neon have?
A million tokens, the same as the other Gemini 4 models.