Skip to header Skip to main navigation Skip to main content Skip to footer
Scott Lawson
Your next upgrade lives here

Google Launches Gemini 4 Neon, a Mid-Tier Model Built for Coding Agents and Long Work

Photograph of the Google campus at the Googleplex headquarters in Mountain View, California.

Google DeepMind has added a new middle option to its Gemini 4 family. Gemini 4 Neon, announced on November 18 in a launch post by Koray Kavukcuoglu, Google's chief AI architect and a senior vice president at Google DeepMind, is positioned between the flagship Carbon model, aimed at longer and more demanding projects, and the faster Hydrogen model. Neon is built for complex coding, analytical reasoning and production enterprise workflows, rather than for a single headline benchmark.

According to Google's model card, Neon is adapted from Gemini 4 Carbon. It accepts text, images, audio and video, and replies in text. Input, reasoning and output share a unified context window of up to one million tokens. A normal synchronous response can run to 64,000 tokens, while a background long decode mode can stretch output up to the full million-token window. The model's knowledge cutoff is March 2026, and Google notes that some domains may be less current than that date suggests.

Pricing is central to the pitch. Through the Gemini API, Neon costs three dollars per million input tokens and fifteen dollars per million output tokens, with output pricing including the tokens used for reasoning. Cached input costs fifteen cents per million tokens. Access is paid only, with no free developer tier listed. Google says that at its default high thinking level, Neon costs 32 percent less per task than Carbon. That comparison needs context, because Carbon is currently on introductory pricing of two dollars input and ten dollars output until November 30, before rising to four and twenty dollars on December 1, when Neon's per-token price becomes the cheaper of the two.

Neon offers four thinking levels, from Low through Medium and High, which is the default, to Extra High, which adds deeper multi-step reasoning. On Google's published figures, the Extra High setting scores 61 on the Artificial Analysis Intelligence Index, ranking third overall behind Claude Fable 5.2 and GPT-6.2 Astra, and ranks first on that service's Coding Agent Index at 58.6. Those are vendor-reported results, and independent teams will need to test the model on their own repositories, documents and budgets before treating the rankings as settled.

The model is rolling out through the Gemini API and Google AI Studio, and it is also available in Android Studio, Gemini for Business and Google Antigravity. Pro and Ultra subscribers can select it in the Gemini app, and GitHub says it is generally available in Copilot Pro+, Max, Business and Enterprise plans. That wide distribution matters, because a mid-tier model succeeds only if developers can reach it inside the tools they already use.

Google is also candid about the limits. The model card warns that Neon can produce fabricated information, says work to improve resistance to jailbreaks is continuing, and notes that the full Safety and Security Report is still to come in the following days, with the current safety results based on automated evaluations rather than human red teaming. For teams considering Neon for coding agents or knowledge work, the sensible next step is a small, fixed-budget trial on real tasks, measuring completed work, review time and cost per task, before any wider rollout.

Artificial Intelligence
Google
Generative AI
Large Language Models
Coding Tools
AI Agents
Technology

Copyright © 2026 Rocky Mountain Madman LLC - All rights reserved

Developed and Designed by Rocky Mountain Madman LLC