Google began opening Gemini 4 Argon to a wider audience on October 13, 2026, making the model available to paying Gemini app subscribers and to developers through the Gemini API. Argon is not new to Google’s lineup: according to Let’s Data Science, Google first showed it on September 30 to a limited group of trusted cybersecurity partners, so this week’s change widens access rather than introducing the model. Reporting in this post comes from Android Headlines and Let’s Data Science, which draw on Google’s own announcements; no wire-service account was available to us, and most performance claims below are Google’s.

Who can use it, and on which plan

In the consumer products, Google AI Pro and AI Ultra subscribers can select Argon in the Gemini app and in AI Mode in Search. Android Headlines lists Pro at $20 a month and Ultra from $99 a month. Google’s agentic development environment, Antigravity, is limited to Ultra subscribers, according to Let’s Data Science.

For developers, the model ID is gemini-4-argon, and the model is offered through the Gemini API and Google AI Studio, per Let’s Data Science. Android Headlines says that using the model in AI Studio requires the top Ultra plan, which conflicts with that account. Because the two outlets disagree, check your own console before building a plan around a particular route.

Pricing, and the November 30 date

Google is charging introductory API rates of $2 per million input tokens and $10 per million output tokens through November 30, 2026. Let’s Data Science says only that standard pricing is higher afterward. Android Headlines puts the post-promotion rates at $4 and $20, double the launch numbers. Treat the second pair as one outlet’s report until Google’s pricing page confirms it.

A worked example shows why the date matters. Take a single request with 100,000 input tokens and 10,000 output tokens. At $2 and $10 per million, that costs $0.20 for input plus $0.10 for output, or $0.30. At $4 and $20 it costs $0.60. These are plain arithmetic on the published per-token rates; they assume flat pricing and ignore any long-prompt tiers or caching discounts, which we have not seen documented for Argon. For a method of turning rate cards into per-task costs, see our worked pricing breakdown for Claude Haiku 5.5.

What the model is for

Google positions Argon as a flagship reasoning model for software engineering, math and science problems, and multi-step agent workflows. The documented context limit is 1,048,576 tokens combined across input and output, and the supported tools include code execution, function calling, file search and URL context, according to Let’s Data Science. Android Headlines describes a 1 million output-token limit; the two descriptions of the limit are not the same, so confirm it in Google’s documentation before relying on long outputs.

The model card reportedly acknowledges hallucinations and occasional slowness or timeouts, and the same outlet advises teams to test their own tasks, tool use, latency and spend before depending on the model in production.

Macro view of a computer processor chip on a dark circuit board with fine gold traces

The claims, labeled as claims

Every item in this section comes from Google through Android Headlines, and none has been independently verified in the sources we read:

  • Argon scored 87% on an embodied-reasoning benchmark and led on Vals Index results for legal, financial and coding tasks.
  • Google engineers reportedly used it inside Antigravity to clean up about six million lines of legacy Chrome code in weeks rather than months, removing a category of browser crashes.
  • On YouTube, Argon reportedly found redundant emoji-rendering code on the backend, and removing it halved comment-server load and reduced latency.
  • In a robotics demonstration, the model watched a video of a person playing cornhole and transferred the motion to Apptronik’s Apollo humanoid, which reportedly sank seven consecutive shots in real-world testing.

These are vendor-supplied case studies. They show what Google chose to publish, not how the model behaves on someone else’s codebase. Let’s Data Science notes that Argon leads or ties on some benchmark entries and trails on others, and that all figures are vendor-reported.

Developer desk with a laptop, mechanical keyboard and notebook in a dim room lit by a monitor

Who gains and who is left out

Android Headlines, in an opinion section clearly separate from its reporting, argues that gating the best coding environment behind a $99-a-month plan hurts solo and hobbyist developers, and that subscribers on a lower tier remain on older models. That is a commentary, not a finding, but the access structure it describes is accurate to the plan details above: the full experience sits at the top tier, with API access available to anyone willing to pay per token.

For teams, the sensible step is cheap and boring: run a fixed set of your own tasks through Argon and your current model before November 30, record cost and error rates, and decide with the higher rate in mind. The promotional window rewards testing, and punishes assuming.

Robotic arm in a bright research lab reaching toward a small wooden block on a white workbench

What to watch next

Watch for three things: Google’s official pricing page confirming the post-November rate, independent benchmark runs that do not rely on Google’s own numbers, and clarification of which plans unlock AI Studio. Competitors tend to respond to pricing moves quickly, so rival rate cards may shift before the promotional period ends.

Sources

  • Android Headlines, Jean Leon, “Google Rolls Out Gemini 4 Argon Worldwide on the App and Developer APIs Today”: androidheadlines.com
  • Let’s Data Science, “Google expands Gemini 4 Argon access across API and subscriber products”: letsdatascience.com
  • Let’s Data Science, “Google Opens Gemini 4 Argon to Gemini App Subscribers”: letsdatascience.com