TellWorldAI World Archive

Home›News›Release

ReleaseGenerative AI for everyone

Google DeepMind releases Gemini 4 Neon

On 18 November 2026 Google DeepMind announced Gemini 4 Neon, rolled it out across the Gemini API and Gemini apps, and released safety and cyber-defense results alongside the benchmarks.

A glowing crystal prism directs a fan of parallel light beams toward a ring of small orbiting cubes.

Background

Gemini 4 Neon is the model in the Gemini 4 lineup that Google positions between Gemini 4 Carbon and Gemini 4 Hydrogen. Google announced it on 18 November 2026 through its DeepMind blog.[1]

What happened

Google DeepMind announced Gemini 4 Neon on 18 November 2026, and it began rolling out the same day in the Gemini API, Google AI Studio, Gemini for business, Google Antigravity and the Gemini app.[1]

Key details

In Google Antigravity, Neon can act as a fast orchestrator that directs parallel Gemini 4 Hydrogen subagents.[1]

On safety-related tests, Google said Neon matched Gemini 4 Carbon with 0.0% cheating on Terminal-Bench 4.0. In its coding cheating evaluation it took shortcuts on 2.45% of tasks.[1] Google also said that with real-time safeguards at its highest thinking level, Neon resisted every attack in Gray Swan's indirect prompt injection benchmark.[1]

Google said Neon ranks second on steerability on the new Agent Arena Index. It reached 51.26% on WebExploitBench v1.2, which Google described as state of the art, at half the cost of competing large models.[1]

Neon for cyber is rolling out to trusted defenders in the Fairwind Program.[1]

Impact

The launch brought the model to developers, business customers, paid Antigravity and Gemini app subscribers, and a set of trusted cyber defenders on the same day.[1]

Terms used in this entry

A benchmark is a fixed set of tasks that scores models on the same test, so results are comparable only when the test and settings match. A thinking level sets how much reasoning a model does before it answers; higher levels usually take longer and cost more per task. A context window is the amount of material a model can consider in one request, measured in tokens, where a token is a small piece of text.

A prompt injection attack hides instructions inside content a model reads, such as a web page or a document, in an attempt to override what the user asked for. A safeguard in this sense is a check that runs while the model works and blocks or flags such attempts. A subagent is a separate model run that a main agent hands a bounded piece of work to, then merges the results from.

Sources

  1. 1

Cite this page: Google DeepMind releases Gemini 4 Neon. TellWorld, https://www.tellworld.net/news/gemini-4-neon-2026. Updated 2026-11-20

Related news

View all