Google rolls out Gemini 4 Argon to paying users, with Ultra-only coding tools

A staged rollout

Google began the wide rollout of Gemini 4 Argon on 13 October 2026, after first presenting it a few weeks earlier with a warning that the company needed to move cautiously toward a full release because of the potential for nefarious use. The model first reached members of Google's Fairwind Program, a group of governments and trusted partners that need access to the company's models with the most advanced cybersecurity capabilities, earlier in October 2026. Developers can now reach it through the Gemini API and Antigravity, Google's agentic software developer platform. Google's benchmarks, published in late September 2026, remain its main evidence of the model's advanced reasoning and agentic coding capabilities. The Gemini app also connects to more than 40 apps, including Instacart, OpenTable, Adobe and Shopify, and Gemini tasks can keep working when the user is away from the computer.

Who can use it

Argon is rolling out in the Gemini app and AI Mode in Search, but only for paying subscribers. Google AI Pro, at $20 a month, and Google AI Ultra both get access. Even those on lower-tier plans will not get full access, according to Ars Technica. In Antigravity and AI Studio, the model is limited to AI Ultra plans. Subscribers on Pro can still generate code with Argon in the Gemini app's Canvas mode, but Ars Technica judges the multi-threaded agentic workflows in Antigravity probably too expensive for them. Google is keeping Gemini 3.1 Pro available for the time being and has declined to say when, or if, Argon will reach free users. The move follows Google's recent restriction of Flash models for free users, which leaves them with Flash-Lite.

Pricing and developer access

Introductory API prices are $2 per million input tokens and $10 per million output tokens. Those prices are valid until 30 November 2026, after which Argon will cost users double those amounts. The Artificial Analysis benchmarking company says Argon matches OpenAI's GPT-6 Astra on its Intelligence Index, a composite score across multiple AI benchmarks, at 60 percent of the cost per task at those introductory prices. Argon can handle up to one million tokens of input and output, which Google says lets users submit whole document archives or large software projects in a single request. Google was supposed to launch a 3.5 Pro model but released Argon instead, according to Engadget.

Internal use and capabilities

Google says thousands of its own employees have used Argon during testing. One example is YouTube's comment system, where Argon identified redundant emoji rendering whose removal cut server load by 45 percent. Engineers also ran Gemini 4 on the proprietary drone physics simulator used in Project Wing, and the AI found critical bugs that had been in the codebase since 2018. In a Chromium test, Argon reorganised about 6 million lines of code within a project of 40 million lines, a job that would have taken months and was completed in a few weeks. In another test, Argon reviewed 134 software bug reports and archived 97 of them as obsolete or duplicate in under three hours. Josh Woodward, VP of Product for Gemini and Google Labs, said that in the Gemini app Argon answers more quickly, with less hedging, repetition and clichés, and can return responses while it is still thinking on longer tasks.

Safety and open questions

Google says Argon has been carefully trained to refuse malicious instructions and includes an automated misalignment monitor that can stop processing if it detects an unintended workflow. The company argues that such safeguards are needed because the model is so capable. Google has not published a date for wider access, so free users remain limited to Flash-Lite for now. Whether the paid-tier restrictions hold once the introductory pricing ends on 30 November 2026 is not yet known.