Google Begins Gemini 4 Argon Frontier Model Rollout

On October 13, 2026, Google announced that Gemini 4 Argon is now fully open to developers, enterprises and consumers, ending the early access and demo phase. The core signal of this move is not the model's capability itself, but the spread of distribution channels: Argon is available through the paid Gemini API, AI Studio, Gemini platform, Google Antigravity, and some Gemini app and AI Mode subscriptions.

For subscribers, Google AI Ultra and Pro subscribers can access the model through the Gemini app and AI Mode in Search. This means Argon covers both the pay-as-you-go developer path and the subscription-based consumer path, with Google trying to establish a unified model entry point between the two types of users.

Google set Argon's entry pricing at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens. This pricing structure clearly favors encouraging cache reuse, and is friendlier to long-task scenarios that need to repeatedly call the same context.

Google says the model supports a unified input-output context window of up to 1 million tokens via the Gemini API, aimed at long-running scientific, coding and agentic tasks. It should be noted that this context window statement comes from Google, and only refers to the support capability via the Gemini API; not all access channels necessarily have the same specification.

Google says Argon ranks first on the Vals Index benchmark, which covers professional work in finance, coding, law and tax. If this ranking holds, it means Argon is highly competitive in professional work scenarios, but this claim comes only from Google and lacks confirmation from an independent benchmarking organization.

According to Google, after internal engineers used Argon through Antigravity, agent-generated code commits increased by more than 35%, and code rollbacks decreased by nearly one third. Google also says Argon helped refactor parts of the Chromium codebase of about 40 million lines, shortening work that would have taken months to weeks.

On the open-source Agent Development Kit, according to Google, changes generated by Argon reduced long-session memory preparation time by 40% and improved cold import performance by 21%. These figures are all Google internal statistics and cannot be externally verified; readers should treat them as vendor self-reports rather than independently verified conclusions.

Google says Argon helped improve quantum error correction software, participated in nuclear fusion mathematics research, and assisted with rare disease genomic analysis and enzyme design. These application cases are all described unilaterally by Google, and the source provides no specific partner institutions or reproducible details.

In robotics, Google says Argon scored 87% on its Embodied Reasoning benchmark, which measures a model's ability to reason about physical and spatial problems in 3D environments. According to Google, in an internal test, Argon watched a video of a person playing a cornhole game, mapped the scene, calculated the required body movements and generated simulation code; after the behavior was transferred to Apptronik's Apollo humanoid robot, it hit seven consecutive shots.

Google calls this process the "Video-to-Sim-to-Real" pipeline, aiming to shorten the creation time of robot training simulations from weeks to hours. The benchmark score and robot test details are not public and cannot be verified, making this a typical vendor self-reported capability demonstration.

While opening the model, Google set up tiered access channels for specific groups. Researchers can apply for free use of the highest Google Antigravity usage tier for six months, subject to eligibility and capacity limits. Google is also giving 10,000 scientists access to the highest Antigravity usage tier for six months, likewise subject to eligibility and capacity limits.

Cybersecurity professionals can apply through Google's Fairwind Program to access Argon under a separate security arrangement. This arrangement indicates that Google has adopted access controls different from the ordinary API for highly sensitive uses, but the source does not explain the specific security mechanisms.

All information in this article comes only from a single source, a report by The AI Insider, with no other independent sources for cross-verification. Among these, the availability channels, pricing, researcher and scientist access programs, and the Fairwind Program are confirmable product launch facts; while the Vals Index ranking, internal coding efficiency data, Embodied Reasoning score and Apollo robot test results are all Google self-reports with no third-party verification.

The source also mentions that Argon added protections against agentic behavior, prompt injection and harmful requests, but the specific mechanisms and effects are not explained in the source. Therefore, regarding Argon's actual capability boundaries and security performance, currently one can only rely on Google's unilateral statements, and readers need to retain cautious judgment.