Key Notes
- Google is rolling out Gemini 4 Argon globally after its limited September preview, with consumer access through the Pro and Ultra subscriptions.
- Introductory API rates run through November 30, while the model shares a million-token budget across input, reasoning and output.
- Published evaluations show leading results on some knowledge-work and coding tests, alongside weaker scores than rivals on others.
Google is rolling out Gemini 4 Argon, its most capable AI model, to developers, businesses and paying consumers worldwide. The October 13 release brings the flagship to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search, while opening paid API access for developers.
This is the wider release of a model Google first unveiled on September 30 through a limited rollout to trusted cybersecurity partners. Our original coverage and accompanying launch brief examined that preview. Today’s announcement makes Argon available for a much broader set of everyday and professional tasks.
Google is positioning Argon around sustained work: complex software engineering, financial and legal analysis, scientific problems and agents that need to complete multiple steps. The launch gives users a chance to test those claims outside the early-access programs that shaped its first public reception.
Who Can Use Gemini 4 Argon?
Google’s rollout announcement separates access by product. Pro and Ultra subscribers can use Argon in the Gemini app and AI Mode. Wider availability therefore does not mean unrestricted free access.
The separate developer announcement points builders to the paid Gemini API in Google AI Studio and Android Studio. It emphasizes code migrations, research and other extended workflows, putting the release inside existing development tools rather than limiting it to a chat interface.
For coding, Google is bringing Argon to the Antigravity desktop app and command-line interface for Ultra subscribers, with access also offered through eligible enterprise subscriptions. Its Antigravity update describes a model intended to handle longer engineering workflows, rather than simply suggest the next few lines of code.
Google says its internal teams saw agentic code submissions rise by 35% over four weeks while post-submit rollbacks fell by 30%. Those are company-reported observations within Google’s development environment, not a guarantee that another organization will see the same gains. The more useful test for a team is whether the agent can complete changes that survive its own review and deployment process.
A Larger Working Window, With Important Limits
The API specifications describe a unified input-and-output window of 1,048,576 tokens. Argon accepts text, images, audio, video and PDFs, but produces text. It supports tools including code execution, function calling, search grounding and preview computer use; native image generation and audio generation are not supported.
The million-token headline needs a qualification. Input, reasoning and output draw on a shared budget, and ordinary requests have smaller generation caps. Google’s developer guide explains that extended generations use a continuation mechanism. Developers should distinguish the model’s overall capacity from how much an individual API request or application actually permits.
Introductory API pricing runs through November 30. The published rates are $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. After that period, Google lists $4 for input and $20 for output. These are usage-based API prices, separate from a Gemini subscription.
Strong Results Do Not Make Every Task a Win
Google’s published model card reports 68.9% on the Vals Index and 77.9% on DeepSWE v1.1. In its comparison, Argon leads GPT-6 Astra and Claude Opus 5.5 on both measures, supporting Google’s case for the model in knowledge work and some demanding software engineering tasks.
The same comparison shows a less uniform picture elsewhere. Argon scores 55.0% on FrontierSWE v2, below Astra’s 65.5% and Opus 5.5’s 62.3%. On Terminal-Bench 4.0, its 57.4% trails Astra at 58.2% and Opus at 66.4%. The results support a competitive flagship, rather than a model that automatically wins every coding workload.
Google’s evaluation methodology also matters. Most Argon results use single-attempt settings and the highest thinking level, while competitor figures generally come from their providers or public leaderboards. Different tasks, tools and reasoning settings can change the comparison. We explored that question in our earlier frontier model analysis; broad access now makes direct testing much easier.
The practical significance of today’s release is that Argon moves from a promising preview into products people can use for their own work. For Google, the next challenge is to turn benchmark strengths into dependable results at a price users accept. For developers and subscribers, the decision comes down to how well the model handles their actual documents, codebases and multi-step tasks.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
