Google Puts Gemini 4 Argon in Front of Paying Users

Google opened its most capable AI model to paying customers this week, ending a two-week run in which Gemini 4 Argon was available only to a small circle of cybersecurity professionals. The October 13 rollout puts the system in the hands of Google AI Pro subscribers at $19.99 a month and Ultra subscribers at $99.99 a month, inside the Gemini app and AI Mode in Google Search, according to Tech Times reporting on the launch.

The release matters because of how long it has been coming. The new system is the first flagship-tier Gemini model Google has shipped above its Flash line in more than seven months, according to model lineage tracking by Artificial Analysis. Google DeepMind's chief AI architect, Koray Kavukcuoglu, announced it on September 30, and chief executive Sundar Pichai repeated the announcement the same day, promising wider availability as soon as possible.

Early access was not for the public at all. The company restricted the model to its Fairwind Program, a limited channel for trusted cybersecurity professionals and select government clients, who received a version without the usual guardrails that block offensive security requests. That phase let researchers test the model on real vulnerability work before consumers ever saw it.

A restricted start before the wider rollout

Google has been feeding the press specific engineering stories to make the case that this is a working tool rather than a demo. The Chrome team used the model to pull core webpage handling out of a roughly six-million-line legacy folder inside the Chromium codebase. Internally, the company says agentic code submissions rose by more than a third while code rollbacks dropped by nearly a third, though those are Google's own figures.

Access is expanding in tiers. Both Pro and Ultra subscribers can now pick the model from an updated model picker in the Gemini app, alongside the existing Flash options. But the features that depend on sustained autonomous work are reserved for the more expensive tier. Google Antigravity, the agentic coding environment built on licensed Windsurf technology, is available only to Ultra subscribers on the higher-priced tier; it lets the model manage, edit and test multiple code files in a single session. Pro subscribers get a lighter Canvas mode for generating code.

The October rollout also introduced a chat-to-tasks toggle in the Gemini app. In tasks mode, the model can work in the background across connected Google services even when the user is not in the app. Gemini now integrates with more than forty third-party services, including Instacart, OpenTable, Adobe and Shopify, according to the company, which Pichai has described at Google I/O 2026 as part of an "agentic transformation" of Google's products.

Benchmarks, with real caveats

Google's own benchmark tables put the model in the lead. The company reports a 77.9 percent score on DeepSWE v1.1, a real-world long-horizon software engineering test, and 68.0 percent on CWE-bench v1, a vulnerability remediation evaluation. DataCamp noted the obvious problem: these are Google's numbers from tests Google selected, and as of early October no outside party had independently replicated them.

The most credible independent evaluation comes from Artificial Analysis, and it complicates the launch narrative. On that firm's Intelligence Index, the model scored 52.56, within half a point of OpenAI's GPT-6 Astra at 52.67, but roughly four to five points below Claude Opus 5.5 at 57.62 and Claude Sonnet 5.5 at 56.00. There is an asterisk on these figures too: the scores reflect the model's high-reasoning configuration, while everyday subscribers use a default medium setting that likely performs differently on reasoning-heavy tasks.

The system does lead the field in one independent metric. Artificial Analysis measured a 15 percent hallucination rate on its Omniscience evaluation, the lowest among all models scoring above 45 on the index. GPT-6 Astra and GPT-6.1 Sol posted hallucination rates of 51 and 54 percent in the same test. The tradeoff shows in accuracy: the model answered 50 percent of questions correctly, compared with 63 percent for Astra.

Pricing for developers tells its own story about the competition. The introductory API rate through the end of 2026 is two dollars per million input tokens and ten dollars per million output tokens, rising to four and twenty dollars respectively in 2027. Independent analysis puts the per-task cost at introductory pricing at roughly two dollars, about three-fifths of Astra's, but that advantage reverses once the promotional rates expire.

The agentic turn raises safety questions

The shift from chatbots that answer questions to systems that take actions on a user's behalf is changing what security researchers worry about. An independent 2026 safety audit of Gemini by Axis Intelligence gave it a composite score of 6.0 out of 10 and identified the expansion of agentic capabilities as a meaningful increase in attack surface.

Google's own product managers seem aware of the tension. The company has added real engineering under the model, including a unified context window of roughly one million tokens shared across input, internal reasoning and output, and three configurable thinking levels that let users trade speed for deeper reasoning. But the context window is not a simple one-million-token output limit: each API request returns roughly sixty-five thousand tokens by default, and reaching the full window requires a continuation mode that chains requests together.

Google has not said whether the model will ever reach the $7.99-a-month Plus tier or free users, only that it intends to reach developers, enterprises and consumers as soon as possible. With Pichai committing between 180 and 190 billion dollars in capital expenditure for 2026 to support AI infrastructure, the company is betting that the demand is there whenever it decides.