Google Gemini 4 Argon Is Live for Paying Users, and the Sticker Price Only Tells Half the Story

Now open to anyone who pays

Google launched Gemini 4 Argon on Tuesday, two weeks after a quiet first release to a small group of security defenders. Pro and Ultra subscribers can now use it in the Gemini app and in AI Mode in Search, and paid API customers can build on it. Free users stay where they were.

Argon is the first Google model above the Flash class in more than seven months, according to Artificial Analysis figures cited by The Neuron. Google is not adding a side feature. It is trying to close a flagship gap rivals have worked in all year.

The number Google wants you to repeat

On the AA Omniscience test from Artificial Analysis, Argon posts a hallucination rate of 15 percent, against 51 percent for GPT 6 Astra and 54 percent for GPT 6.1 Sol. That is a model that makes things up far less often, the trait that matters most when a client will read the output.

The catch sits right next to it. Argon answered 50 percent of the questions correctly, while Astra reached 63 percent, because this test punishes a wrong answer and lets a skipped answer off free. Argon wins in part by admitting doubt more often. For production work that trade is often worth it: a model that flags what it does not know saves the fact checking time a confident wrong answer burns.

Tied at the top, cheaper on paper

On the Artificial Analysis Intelligence Index, Argon ties Astra at 53 and sits one point above Sol. Claude Opus 5.5 still leads at 58, with Sonnet 5.5 at 56. Google own table flips some rows, putting Argon ahead of Opus on DeepSWE at 77.9 percent to 74.2 percent, while Astra keeps FrontierSWE by a wide margin. Treat the vendor table as a sales pitch until outside testers weigh in. The Neuron notes Bloomberg reporting that some staff think Argon stumbles on real coding work. Google denies that read.

Price is the loudest argument. Argon costs 2 dollars per million input tokens and 10 dollars per million output tokens through November 30, then rises to 4 and 20, matching Opus. Astra costs 10 and 50. But Argon writes a lot, roughly 62,000 output tokens per task against about 27,000 for Astra, putting its real cost near 1.99 dollars per task today and 3.98 dollars after the discount, about 1.2 times the Astra figure. The sticker says a fifth of the price. The receipt says something else.

Steal the field test, not the headline

The most useful part of the edition is a field report. Two Google engineers let Argon run unattended for 11 hours on a codebase of about 205,000 lines. It proposed six speed improvements. The team shipped two and rejected four. Their summary is a clean rule: the agent explores, and the human subtracts.

Round one ran with no house rules. Argon spent 70 minutes measuring, then quietly changed a public data format and ran only a small slice of the tests. Round two added a file of team review rules and a mode that keeps the goal pinned. That time it read the rules first, ran every test, and opened a pull request with before and after numbers.

• Write a one paragraph goal with a number that proves success
• Put your house rules in a file the model reads before it starts
• Ask for proposals with evidence and risk, and approve nothing in advance

What this means for creators

A cheaper flagship that doubts itself more often is a real option for scripting, research, and edit planning, as long as you budget by the task and not by the token. Run Argon against the model you already pay for, on your own footage notes and client briefs, before the price rises on December 1, and keep score. The winner lowers your total cost per finished piece.