What A Time To Be Alive
Choose language: English

Google releases Gemini 4 Neon: close to top level, faster and cheaper

Google DeepMind is releasing Gemini 4 Neon, a fast AI model for programming, analysis and jobs that take several steps. It comes close to the top model Carbon in power, but costs less. Subscribers to Google AI Pro and Ultra can now choose it in the Gemini app.

Title card with the words Gemini 4 in white and Neon in light blue, the Google logo in the upper right and a dark planet with a blue glow in the lower right against a starry sky.

Within the Gemini 4 family, Gemini 4 Neon sits between the fast, cheap Hydrogen for everyday tasks and Carbon, the heaviest model for the hardest jobs. According to Google, Neon almost reaches Carbon's level, but works faster and costs less. It is meant for complicated programming work, analyses and tasks that consist of several steps.

The model can handle text, images, video, sound and PDFs. In one go it can take in about a million tokens of information: the pieces of text an AI model reads and writes in. That is roughly a few thick books. Users can choose how long the model thinks, in four settings from Low to Extra High.

Where and for how much

In the Gemini app, Neon replaces the models Gemini 3.1 Pro and Argon for subscribers to Google AI Pro and Ultra. Developers can use it through the paid Gemini API, the connection their own programs use to call the model, in Google AI Studio. Companies can use it through Gemini for business and the programming platform Google Antigravity.

Through the API, Neon costs 3 dollars per million tokens of input and 15 dollars per million tokens of output. That is twice as expensive as Hydrogen. From 1 December, Carbon costs 4 and 20 dollars, so Neon is a quarter cheaper. Up to and including 30 November, Carbon and Argon still have a temporary price of 2 and 10 dollars.

How good is it?

The independent testing firm Artificial Analysis gives Neon 61 points on its Intelligence Index in the highest thinking setting, a combined score from a series of tests. Claude Opus 5.5 from Anthropic scores 59 there and GPT-6.2 Sol from OpenAI 57. Only Claude Fable 5.2 (65) and GPT-6.2 Astra (62) score higher; according to Google, those are larger, slower models. In Google's own comparison, Neon wins 16 of the 19 tests; Opus 5.5 is better on two programming tests and Sol on a test for reading charts.

A separate version, Neon for cyber, will go to a select group of security specialists in Google's Fairwind Program in the coming days. They can use it to find, check and close weak spots in software.

What does this mean for the road to the singularity?

Models that program on their own and carry out long jobs already exist. With such jobs, what counts is not only how clever a model is, but also what each intermediate step costs: a model that takes a hundred steps pays a hundred times. Neon shows that near-top level is becoming cheaper and cheaper, three weeks after Carbon appeared.

That puts pressure on everyone's prices. If Google offers near-top quality for a quarter less, Anthropic and OpenAI have to follow or offer something Google does not have. For users that means: what only a large company can afford today, a small team or a freelancer will be able to have done as well, such as having all the code of a program checked or hundreds of documents summarised. The cheaper those steps become, the more often a model can check and improve its own work before a person sees it.

What is the technological singularity? Read the explanation →

Sources

  1. 9to5Google
  2. Google
  3. OfficeChai
  4. Artificial Analysis