Gemini 4 Argon

Google announced Gemini 4 Argon in a post dated 30 September 2026 from one of the senior leaders of Google DeepMind. The company calls it a frontier model built for long-horizon work in software engineering, for enterprise knowledge work such as legal and finance, and for cybersecurity defense, and says it already runs in Google's internal workflows.

Who can use it now

For now, access is limited to a group of trusted cyber defenders in Google's Fairwind Program. Google says it is taking part in the US government's voluntary process for pre-release model access and will widen access gradually, then reach developers, enterprises and consumers, with Google AI Ultra subscribers and paying API customers going first. It gave no date. The Verge reported the release independently on the same day.

Pricing and specifications

Google gives an introductory price per million tokens, with cached input 95 percent cheaper, and a higher rate that applies after the introductory period. Google's post gives no model identifier, context window or knowledge cutoff, and when the Gemini API's model, pricing and changelog pages were read on 1 October they did not mention the model. Artificial Analysis lists a context window of 1 million tokens, with text and image input and text output.

Safeguards Google describes

Google's post lays out protections against cyber and CBRN misuse, monitoring of the model's reasoning and actions for misalignment, and strong results on Gray Swan's indirect prompt-injection benchmark. Google's evaluation figures are its own claims.

Independent ranking

Independent testing by Artificial Analysis puts it at 53 on the Intelligence Index (version 4.3.2) at the High setting, eighth among 223 models. Within that version GPT-6.1 Sol scores 52 at maximum effort and Claude Opus 5.5 scores 58.