On 5 October, we showed you a model that almost nobody could use: Gemini 4 Argon, the most powerful Google has ever built, initially reserved for cybersecurity experts.
Since Tuesday, it has arrived in the Gemini app. A chance to answer the only question that matters: how good is it really?
Who can use it
Paying Gemini subscribers. Argon is available in the app for the AI Pro and AI Ultra plans, as well as in Antigravity, Google's coding tool.
Not free users. Since 9 October, the free version of Gemini has been limited to Google's smallest model, as explained in our article on free Gemini. Nothing has changed on that front.
Not all developers yet. Google promises to widen access "as soon as possible", without a date. The price, however, is known: 2 dollars per million input tokens and 10 dollars for output during a launch period, then 4 and 20 dollars.
The version without cybersecurity guardrails, meanwhile, remains reserved for defenders in the Fairwind programme.
What Google promises
According to Google, Argon answers more directly, with fewer hedging caveats, repeats itself less and avoids stock phrases. On long tasks, it keeps the user informed of its progress more often.
It can also produce gigantic responses, up to a million tokens at once, against 64,000 for the previous generation: enough to write the equivalent of several books in a single response. And Gemini now connects to more than 40 applications, from restaurant booking to creative software, to carry out tasks on your behalf.
What independent tests measure
Google's figures are flattering, as always. Here is what two independent organisations, Artificial Analysis and Vals, which test models with their own methods, have to say.
| Measure | Gemini 4 Argon | Claude Opus 5.5 | GPT-6 Astra | GPT-6.1 Sol |
|---|---|---|---|---|
| Overall index (Artificial Analysis) | 53 | 58 | 53 | 52 |
| Hallucination rate (Artificial Analysis) | 15% | not published | 51% | 54% |
| Vals ranking (41 models) | 1st, 68.9% | 67.0% | not published | 61.2% |
| Terminal coding | 57% | 60% | 59% | not published |
| Average cost per task | $1.99 | $5.98 | $3.26 | cheaper |
Its real strengths
It makes things up far less. This is the most striking result. On the Artificial Analysis test, Argon invents an answer in 15% of cases where it does not know, the lowest rate measured among large models. Its direct rivals from OpenAI exceed 50%.
It excels at demanding office work. Analysis of long documents, legal and financial tasks, assignments that stretch over many stages: that is where it takes the lead, both in Google's tests and in Vals's independent ranking.
It remains reasonably priced. Even though it produces a lot of text, an average task costs it less than competing high-end models.
Artificial Analysis's test puts detailed knowledge questions to models, across many domains. It does not simply count correct answers: it looks at what the model does when it does not know. Saying "I don't know" does not lose points. Inventing a wrong answer does. A cautious and honest model can therefore do better than a model that knows more things but fills its gaps by making them up. That is exactly the ground on which Argon stands out.
Its limits
It is not the strongest everywhere. On Artificial Analysis's general index, it matches GPT-6 Astra, but remains behind Claude Opus 5.5, and even behind Claude Sonnet 5.5, at 56.
Pure programming is not its strong point. On development tasks carried out in a terminal, it is beaten by Anthropic's and OpenAI's models.
It is talkative. Argon produces a lot of text to reach its ends. As a result, even at the same price per token as GPT-6.1 Sol, a task costs it noticeably more, around 2.7 times more according to Artificial Analysis.
Who is it worth it for?
If you already pay for a Gemini subscription and you work on long documents, research, legal or financial files, try it: that is precisely where it shines, and its low rate of invention is a real asset.
If you are mainly looking for a programming tool, or if you count every penny, other models remain better placed. And if you use Gemini for free, you will not see it.
What we take away
Gemini 4 Argon does not dethrone everyone. But it wins on ground that matters more and more: reliability. A model that invents three times less than its direct competitors may be more useful day to day than one extra point in a ranking.
That promise nevertheless hides another, more delicate one: Google touts a model that "hesitates less". Can one hesitate less and invent less at the same time? We devote our reflection this evening to that.