Gemini 4 Argon: 13 Wins, 5 Losses (Google's Own Table)

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Gemini 4 Argon: 13 Wins, 5 Losses (Google’s Own Table) · Channel: Prompt Engineer 48 · 05:42 · Watch on YouTube

Topics: AI evaluation and benchmarks, Cybersecurity and software security, Software engineering with AI, AI models, AI safety

Summary

The video reviews Google DeepMind’s newly announced Gemini 4-Argon using Google’s benchmark table, noting 13 wins, one tie, and five losses across 19 rows. It highlights reported strengths in knowledge work, coding, long-context tasks, vision, and cybersecurity, alongside Google’s claims about its internal use of the model. The video also covers safety measures, introductory pricing, access restrictions, and benchmarks where other models perform better. The presenter emphasizes that they have not tested Argon firsthand and recommends waiting to evaluate it on your own work, particularly for terminal-style coding.

Key points

Gemini 4-Argon was announced on September 30 and is initially available only to trusted testers through the Fairwind program.
In Google’s table of 19 benchmark rows, Argon has 13 outright wins, one tie, and five losses.
The video highlights results in knowledge work, including finance and legal benchmarks, and coding benchmarks such as Deep SWE and Vibe Code Bench.
Google says it used Argon to migrate over 800,000 lines of C and C++ to Rust in the Fuchsia Zircon kernel.
Google reports that Argon performed well on long-context tasks, including the 256,000-to-1-million-token row, and that its output limit is 1 million tokens.
On cybersecurity tests, Argon tied for first on CWE Bench and scored above Gemini 3.8 Flash Cyber on Google’s vulnerability-finding test.
Google reports a 0.7% attack success rate for Argon on Gray Swan’s prompt-injection test, where lower is better.
Introductory pricing is $2 per million input tokens and $10 per million output tokens; the stated later rates are $4 and $20, respectively.
The presenter notes that the figures come from Google’s own table and that they have not independently tested the model.
The video advises people focused on terminal-style coding agents to wait and test Argon on their own work.

Chapters

Start	Chapter	Summary
00:00	Announcement and benchmark overview	The presenter introduces Gemini 4-Argon, its limited availability, and the plan to review its benchmark results.
01:00	Knowledge work, coding, and Google’s internal use	The video covers Argon’s reported benchmark wins and Google’s claims about using it for code migration and other internal tasks.
02:36	Long context, vision, and cybersecurity	The presenter discusses long-context and vision results before reviewing cybersecurity benchmarks and a vulnerability Google says it found.
03:39	Safety and pricing	The video describes Argon’s prompt-injection test result and listed safeguards, then gives its introductory and subsequent prices.
04:41	Where Argon loses and who should watch	The presenter lists the five benchmark rows where other models lead, explains the limits of the evidence, and discusses access and potential users.

Quotes

that’s 13 outright wins, one tie, and five rows where somebody else is ahead. — 01:00

the output limit going from 64,000 tokens to 1 million tokens. — 02:36

Argon is at 0.7% attack success — 03:39

introductory pricing is $2 per million input tokens and $10 per million output tokens. — 04:09

I haven’t run this model, and you can’t either right now. — 04:41

Concepts and entities

Gemini 4 Argon, Google DeepMind, Fairwind Program, GPT-6 Astra, Claude Opus 5.5, CWE Bench, Gray Swan, Wiz

Full transcript (original language)

00:00 All right. So, Google just dropped Gemini 4-Argon, and you can see this is a really big table. Now, mind you, nobody outside a small group can use this model yet, so I can’t test it for you, but I’ve gone through the whole announcement and the whole benchmark table. And in this video, I’ll show you what looks really good and where it actually loses. So, what is it? It’s the new frontier model from Google DeepMind announced on the 30th of September. They say it’s built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense. And it’s rolling out today only to a set of

00:30 trusted testers through something they call the Fairwind program. So, here’s what we’ll do. First, the table against GPT-6, Astra, Claude Fable 5.1, and Claude Opus 5.5. Then, what Google is already doing with it inside Google. Then, cybersecurity, safety, and the price. And at the end, the rows where it doesn’t win because that matters, too. So, let’s go ahead and look at the table. There are 19 rows here. Gemini 4-Argon is in the first column, and the blue cells are where it’s number one. If you count them, that’s 13 outright wins,

01:00 one tie, and five rows where somebody else is ahead. Starting with knowledge work, Vells Index 68.9%, and the next best is Opus 5.5 at 67. Automation bench, which comes from Zapier, 51.3, and the next one is Opus at 42.5. Vells finance agent 65.4 against 58.9. And look at Harvey’s legal benchmark, 19.6%, where the others are between 3.8 and 6.7. Now, coding, Deep SWE version 1.1 77.9%. That’s the number

01:31 Google calls state-of-the-art on long-horizon real-world tasks. GPT-6, Astra, and Opus 5.5 are both around 74. Vibe code bench, 91.9, and Google says it’s good at debugging, big code base migrations, and algorithm design. Here’s the part I found more interesting. Google is already using it on its own code, is migrating C and C++ to Rust across their infrastructure, over 800,000 lines in the Fuchsia Zircon kernel, and on the lib gav one video decoder they say the result is 2.7 times faster than the previous rust port. And

02:03 two more numbers from inside Google in quantum computing it optimized an algorithm space-time resources 40% better than the published baseline. And in a fleet-wide memory analysis it freed up over 300 TB bytes with an estimated 500 TB bytes to 1 PB byte in total savings. Again, these are Google’s numbers, not mine. Next long context graph walks up to 128,000 tokens 99.7% but look at the second row 256,000 up to 1 million tokens. Argon gets 84.2 and the next best is 71.8. That’s a big

02:36 gap and it matches the other change which is the output limit going from 64,000 tokens to 1 million tokens. On vision and computer use LV bench which is long video understanding 91.7% photography 71.6 and agents last exam 39.5% which is the top number in that row too. Now cybersecurity which is what this release is really about on CWE bench it’s tied for first place at 68% along with GPT 6 Astra and Grok 4.7. On Google’s own test finding vulnerabilities across 20 programming

03:07 languages it gets 85.8% compared to 71 for Gemini 3.8 flash cyber. And on the Wiz penetration test 70.9 against 58.2. And working with Wiz Google says it found a critical vulnerability in healthcare software that previous models missed. So it’s not only finding bugs, it’s validating them and then patching them. You can see that’s the whole point of the fairwind program. Now safety and this chart is really something. This is gray swan’s prompt injection test and lower is better. Argon is at 0.7% attack

03:39 success, Claude Opus 5.5 and Fable 5.1 are at 1%. GPT 6 Astra is 8.5 and at the far end Grok and Kimmy are both over 50. Google calls Argon its most resilient model yet against indirect prompt injections. They also list four things they’ve done, refusing harmful cyber and CBRN requests with activation monitoring, defense against prompt injection, monitoring the chain of thought and the actions, and stopping the run if it goes beyond what you asked, and sandboxing the environment

04:09 before the high-risk evals. Now, the price, since that’s what I care about, introductory pricing is $2 per million input tokens and $10 per million output tokens. After that, it goes to $4 for input and $20 for output, and cached input is 95% off the standard rate. Okay, so now where it loses, Frontier Swivel Argon is at 55% and GPT-6 Astra is at 65.5. Terminal Bench 4.0 Opus 5.5 wins with 66.4. Post-train Bench Opus

04:41 again, Terminal Bench Science Astra is at 68.1 against 57.6 and OSWorld 2.0 72.6 against 69.2. So, that’s five rows. And two honest things. First, these are Google’s own numbers from their own table, and the methodology is on their DeepMind site. Second, I haven’t run this model, and you can’t either right now. Access is only for trusted cyber defenders. They say a US government pre-release process is underway, and API customers and Google AI Ultra

05:11 subscribers get it first when it opens up. So, who is this for? If you do big code migrations, long documents, finance or legal work, or you run security for a company, this one is worth watching. If you mostly use terminal-style coding agents, the table says wait and test it on your own work. So, in this video, we’ve seen what Gemini for Argon is, how it does against GPT-6 Astra and the Claude models, and where it loses. As soon as it opens up, I’ll test it myself and show you the real results in the next video. Let me know which benchmark you want me to try first. I’ll see you

05:42 in the next one. Thank you. >> [bell]