Gemini 4 Argon ties GPT-6 Astra for 60% of the cost. Read the footnote on that price.
Google's new frontier model matches OpenAI's flagship on Artificial Analysis' index. The cheaper bill depends on a 50% launch discount, and Argon writes more than twice as many tokens per task.
The number that matters in Google’s Gemini 4 Argon launch is 62,000. That’s roughly how many output tokens Argon produced, on average, for each task in Artificial Analysis’ Intelligence Index. GPT-6 Astra, the OpenAI model Argon is being measured against, used about 27,000.
Keep that figure in your head, because most of the headlines about Argon are about price, and price is where it bites.
What Google actually announced
Google published the Argon post on September 30, under the name of Koray Kavukcuoglu, who runs Google DeepMind’s model work. Argon is the first Gemini 4 model, and Google calls it its new frontier model. You can’t use it yet. It’s going first to a group of “trusted cyber defenders” through Google’s Fairwind Program, plus Google’s own teams. Paid API customers and Google AI Ultra subscribers come next, with no date given.
The pitch leans hard on security. Google says Argon “can autonomously find, validate, and patch critical software vulnerabilities,” and that for vetted defenders it will ship “without cyber guardrails.” It also says the model found a critical flaw exposing personal data in healthcare software used by hospitals worldwide. SecurityWeek pointed out that Google didn’t name the software or say whether it’s been fixed. I’d like to know both before I treat that as more than a demo story.
There’s a lot more in the post. A 1 million token output limit, up from 64,000. A claimed 77.9% on DeepSWE v1.1. A 68% tie for first on CWE-bench v1, alongside GPT-6 Astra and Grok 4.7. Internal examples too, like Argon agents freeing more than 300 TiB of memory across Google’s data centers. All of those are Google’s own numbers.
The 60% figure has a clock on it
The independent check came from Artificial Analysis, which benchmarks models for a living. Its verdict was friendly: Argon scores 53 on its Intelligence Index, level with GPT-6 Astra and one point above GPT-6.1 Sol, and 23 points above Gemini 3.1 Pro Preview. After months of Google looking a step behind, that’s a real result.
Then the cost line. At launch prices Argon costs $1.99 per index task, which is about 60% of Astra’s $3.26. That’s the stat getting passed around.
It depends on a promotion. Argon’s list price is $4 per million input tokens and $20 per million output. Google is running it at half that, $2 and $10, as an “introductory price.” Artificial Analysis says the discount lasts at least a month and that Google hasn’t confirmed an end date. When it ends, Argon’s cost per task goes to $3.98, about 1.2 times Astra’s.
| Model | Intelligence Index | Cost per index task |
|---|---|---|
| Gemini 4 Argon, launch promo | 53 | $1.99 |
| Gemini 4 Argon, list price | 53 | $3.98 |
| GPT-6 Astra (max) | 53 | $3.26 |
| GPT-6.1 Sol (max) | 52 | about $0.74 (Argon promo is 2.7x Sol) |
And that’s where the 62,000 comes back. Artificial Analysis says plainly that Argon’s lower cost comes from cheaper tokens, “rather than reduced token use.” It thinks longer and writes more. With a 1M output cap, it can write a lot more. If you’re paying per token, a model that talks twice as much needs to cost less than half as much per token just to break even. At list price it doesn’t.
The Sol figure in the table is worked out from Artificial Analysis' "2.7x" comparison, not quoted directly. Check the Artificial Analysis model page for the current numbers before you budget anything.
The hallucination number is good news, with a catch
The other stat being shared is Argon’s 15% hallucination rate on Artificial Analysis’ AA-Omniscience test, against 51% for GPT-6 Astra and 54% for GPT-6.1 Sol. That’s a big gap.
The catch is in how the test works. The hallucination rate looks at answers that weren’t correct and asks how many were confident wrong guesses instead of “I don’t know.” Argon is much more willing to say it doesn’t know. Its accuracy on the same test is 50%, which is 13 points under Astra’s 63% and 5 points under Google’s older Gemini 3.1 Pro Preview. The overall score comes out at 42 for Argon and 43 for Astra.
In plain terms, Argon guesses less, but it doesn’t get more answers right. For a security tool that might be exactly what you want, since a scanner that invents a vulnerability wastes a team’s afternoon. For a research assistant, a lot of “I don’t know” gets old fast.
A real comeback on a temporary price
I think Argon is a genuine comeback for Google, and the independent index backs that up. I’m less sold on the “60% of the cost” framing, because it describes a price Google has openly said is temporary, on a model almost nobody can buy yet. By the time Argon reaches regular API customers, the promo could be gone and the token habit will still be there.
We just watched OpenAI cut prices in half in a single week, in the price fight that followed Claude Opus 5.5. Google launching at half price looks like the same fight from a different corner. If you’re picking a model for real work, test it on your own tasks and count the output tokens, not just the rate card.
The things to watch are the end date of the discount, and whether Google names the hospital software once it’s patched.
Sources
- Google: Gemini 4 Argon: our next era of frontier intelligence (September 30, 2026)
- Artificial Analysis: Gemini 4 Argon: Google is back as one of the top three labs in intelligence achieved (September 30, 2026)
- SecurityWeek: Google Launches Gemini 4 Argon With Guardrail-Free Access for Vetted Defenders (October 1, 2026)
- TechCrunch: Google releases Gemini 4 Argon, called its most powerful model yet (September 30, 2026)