Gemini 4 Neon accelerates enterprise code at Google

Google deploys Gemini 4 Neon, a model designed for complex software development and agent tasks. The company positions it between Hydrogen and Carbon, with a lower API price and a one million token context.

Google deploys Gemini 4 Neon, a model designed for complex software engineering tasks and agent workflows, aiming to cost less than heavier models. The launch primarily targets enterprise developers, as AI purchases increasingly shift to pay-as-you-go and cost-per-task models. Neon sits between Gemini 4 Hydrogen, designed for speed, and Gemini 4 Carbon, reserved for the most demanding work.

Un modèle intermédiaire qui vise le quotidien des équipes techniques

Neon is not presented as a single replacement but as an additional component in the Gemini 4 family. Google describes it as capable of working in large code repositories, fixing multiple files at once, and refactoring beyond superficial adjustments. In Antigravity, its development environment, it can also coordinate several Gemini 4 Hydrogen sub-agents in parallel.

This assembly logic is as important as raw performance. In enterprises, software projects are not just a single neatly packaged request; they accumulate dependencies, legacy files, and production constraints. Google thus aims to position Neon as a model able to handle the load without exploding costs. The kind of compromise that makes purchasing teams smile as much as it interests developers.

Des scores élevés, mais la facture reste au centre du jeu

On benchmarks cited by Google, Neon leads the Artificial Analysis Coding Agent Index with a score of 58.6 and reaches 65.2% on Terminal-Bench 4.1, a test evaluating AI agents in terminal environments. The company also states that Neon costs 32% less per task than Gemini 4 Carbon at the default high reasoning setting in the Artificial Analysis Intelligence Index evaluations. On some competing rankings, Google reports savings of 50% to 80% compared to leading models on Vals Index and Agent Arena [to be verified according to exact scopes].

The announced API price is 3 dollars per million input tokens and 15 dollars per million output tokens. Cached inputs cost 0.15 dollars per million tokens, and the model accepts a context window of up to one million tokens. In other words, it is designed for long tasks with a lot of history to keep in mind, reducing back-and-forth but potentially increasing consumption quickly if usage is not finely managed.

Pourquoi Google pousse ce modèle maintenant

Google's strategy fits a market where companies no longer want just the best benchmark score. They seek controlled total cost, good integration with existing tools, and code quality that holds in production. Anand Joshi, managing director at JP Data, points out that long context windows can cause token estimates to spiral, pushing organizations toward smaller, more efficient models if quality follows.

Prabhu Ram, vice president for industry research at CyberMedia Research, believes companies might adopt a tiered approach rather than relying on a single premium model. His point is simple: you don’t use the same engine to fix a line of code, relaunch an autonomous agent, or rewrite a critical service. In this context, Neon becomes an intermediate option for tasks where Carbon would be too costly and Hydrogen too limited.

La production reste le vrai test

Despite progress, production deployment remains the friction point. Joshi recalls that AI-generated code can resemble a black box to developers, complicating diagnosis when an application breaks under real conditions. Without clear understanding of what the agent modified, companies hesitate to go beyond prototyping.

Pareekh Jain, CEO of Pareekh Consulting, adds that orchestrating multiple agents can create modification conflicts and coordination errors, especially on long tasks. Maintaining reliability in critical business environments remains a central issue. In this context, the most plausible short-term idea remains supervised autonomy, where AI performs more work while humans retain final control.

Key points

  • Neon is positioned between Hydrogen and Carbon in Gemini 4.
  • Google announces 58.6 on the Artificial Analysis Coding Agent Index.
  • The model accepts up to one million tokens of context.
  • API: 3 dollars input, 15 dollars output per million tokens.
  • Google makes it available in Gemini API and Antigravity.
  • Terminal-Bench 4.1: 65.2% according to Google.

By the numbers

  • 32% lower cost per task — compared to Gemini 4 Carbon, according to Google and Artificial Analysis.
  • 1 million tokens — maximum context window announced by Google.
  • 3 dollars — price per million input tokens, Neon API.
  • 15 dollars — price per million output tokens, Neon API.
  • 65.2% — score shown on Terminal-Bench 4.1.

To read