Gemini 4 Argon: what Google claims, and what Vals AI and Wiz show
Google released Gemini 4 Argon on September 30th. It's built for long jobs: coding that runs for hours, legal and finance work, and cyber defense. Google's own post makes the claims. Two companies outside Google, Vals AI and Wiz, show how some of those claims look from the outside.
One answer can now run to 1 million tokens, up from 64,000
Start with the change that's easiest to miss. Argon's output limit rises to one million tokens, up from 64 thousand. That is room for a long report or a large code change in a single answer, instead of one stitched together across many turns.
On Google's long coding test, Argon leads at 77.9%
On DeepSWE, a test of long software engineering tasks, Google's chart puts Argon at about 78 percent. The closest, Claude Opus 5.5 and GPT-6 Astra, sit at about 74. Google calls that a new state of the art. It is Google's number, from Google's own run.
Google's own table shows where Argon trails
Google also published a full results table. At the top, Argon leads on most of the knowledge work rows. Scroll down to coding, and the picture is mixed. On FrontierSWE, Argon scores 55 percent, and GPT-6 Astra about 66. On Terminal-bench, Claude Opus 5.5 leads at about 66, with Argon at 57. Google put these rows on its own page.
Vals AI's own index ranks Argon first, at 68.9%
Vals AI runs its own index of finance, coding, legal and tax tasks, weighted by each field's share of the US economy. As of October 2nd, it ranks Argon first at about 69 percent, just ahead of Claude Sonnet 5.5 and Claude Opus 5.5 at about 67. That is the same score Google quotes in its own table.
Wiz runs Argon to find exposed systems in hospitals and public services
Wiz, a cloud security company, runs a program called Scan for Good for hospitals, public services and nonprofits. Its AI scanning step is powered by Gemini 4 Argon. Wiz reports 475 high or critical findings so far, across about 330 thousand public endpoints. Wiz Research checks each one and discloses it privately to the organization affected.
Approved cyber defenders get it first, at $2 in and $10 out per million tokens
The same post sets the price. It starts at 2 dollars per million input tokens and 10 dollars per million output tokens, then rises to 4 and 20. For now, only approved partners in Google's Fairwind Program can use it, for internal security work. Paid API customers and Google AI Ultra subscribers come next, with no date given.
Argon launches at Claude Sonnet 5.5's price per token, then rises to Claude Opus 5.5's
| Model | Input | Output |
|---|---|---|
| Gemini 4 Argon introductory | $2 | $10 |
| Claude Sonnet 5.5 | $2 | $10 |
| GPT-6.1 Sol | $2 | $10 |
| Gemini 4 Argon after the introductory period | $4 | $20 |
| Claude Opus 5.5 | $4 | $20 |
| GPT-6 Astra | $10 | $50 |
| Claude Fable 5.1 | $10 | $50 |
Standard API list prices in U.S. dollars per 1 million tokens, before caching and batch discounts. GPT-6 models charge more for long prompts. Argon's prices are announced in Google's post and not yet listed in the Gemini API. Checked 5 October 2026 on Google's post and Anthropic's and OpenAI's pricing pages.U.S. dollars per 1 million tokens, before caching and batch discounts. Argon's are from Google's post, not yet the Gemini API. Checked 5 October 2026.
How does that compare? Argon's introductory price, 2 dollars in and 10 dollars out per million tokens, is what Claude Sonnet 5.5 and GPT-6.1 Sol charge today. Its later price, 4 and 20, is the same as Claude Opus 5.5. GPT-6 Astra and Claude Fable 5.1 list at 10 and 50, five times Argon's starting price. These are list prices per token, before discounts for caching and batch jobs.
Artificial Analysis scores Argon level with GPT-6 Astra, at $1.99 a task to Astra's $3.26
As tested: Argon high, the others max. Argon costed at its $2/$10 introductory price.
Artificial Analysis ↗Artificial Analysis runs its own index of tests on each model. Argon scores 53, level with GPT-6 Astra and Claude Fable 5.1, and behind Claude Opus 5.5 at 58. It also prices each run. At its introductory price, Argon averaged about 2 dollars a task, against about 3 for GPT-6 Astra and 6 for Claude Opus 5.5. Each model ran at its own setting, and Argon's figure would roughly double at its later price.
Not in the Gemini API yet: no model ID, no API price listing
Argon was missing from Google's Gemini API models page on October 4.
One thing to watch. Four days after launch, Argon still isn't on the Gemini API models page, and it has no model ID. When it shows up there, developers can try it themselves. Until then, the benchmark claims are Google's, and Vals AI, Artificial Analysis and Wiz give the outside view. Every page shown here is linked below the video.
