Gemini 4 Argon Is Google’s Most Powerful AI Model + Early Tests!

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 Is Google’s Most Powerful AI Model + Early Tests! · Channel: WorldofAI · 14:08 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, AI models, Multimodal models

Summary

The video presents Gemini 4 Argon as a newly announced Google model, currently available only to selected cybersecurity partners and trusted testers, with a broader rollout expected later. The host highlights its reported long output limit, pricing, benchmark results, and low hallucination rate, while noting that it may not lead in coding or computer-use tasks. Early demonstrations show it generating detailed 3D scenes, simulations, and SVGs, though some examples use earlier Gemini checkpoints rather than confirmed Argon builds. The video concludes that Argon could be a strong daily-use model and a significant step in Google’s effort to regain ground in the AI race.

Key points

Argon is initially limited to selected cybersecurity partners and trusted testers through Google’s Fairwind program; paid API customers and Google AI Ultra subscribers are expected to get access later.
The host reports introductory pricing of $2 per million input tokens and $10 per million output tokens, with cached input priced 95% lower.
The model is said to support up to 1 million output tokens, aimed at long-running agentic and coding tasks.
The video cites strong results on several benchmarks, including 77.9% on Deep Sway 1.1, 51.3% on Automation Bench, and 91.7% on LV Bench.
The host says Argon scores 53 on the Artificial Analysis Intelligence Index and has a reported 15% hallucination rate.
The host describes Argon as competitive in cost and performance, but not necessarily the cheapest frontier model or the best in every domain.
Demonstrations include a detailed voxel sanctuary, a PS5 controller SVG, a simulated float plane, Halloween characters, a mechanical bee, and a rendered Colosseum.
Several showcased examples are described as earlier Gemini checkpoints, so the video does not confirm that all demonstrations are results from the final Argon model.
The host views Argon as a potential daily driver, especially for multimodal tasks, while saying it may lag behind other models in coding and web development.

Chapters

Start	Chapter	Summary
00:00	Argon announcement, access, and pricing	The host introduces Gemini 4 Argon, describes its limited initial release, and highlights its reported pricing and one-million-token output limit.
01:36	Benchmarks and cost efficiency	The video reviews reported benchmark results, compares Argon’s performance and cost with other models, and emphasizes its low hallucination rate.
05:16	Availability and early demonstrations	The host discusses the model’s reported scale and rollout plans, then showcases examples including a voxel scene, an SVG, and a 3D physics simulation.
08:24	Interactive and 3D generation examples	Further demonstrations show a mechanical bee, a Colosseum with day-and-night cycles, and a side-by-side 3D world-building comparison.
11:33	Assessment and conclusion	The host weighs Argon’s potential as a daily driver, discusses Google’s AI progress and the limits of claims about ASI, and closes with a positive assessment.

Quotes

this may be the best cost-to-performance model that we have ever seen. — 00:31

Google has increased the maximum output from 64K to 1 million tokens output. — 01:05

Argon has a 15% hallucination rate — 04:14

it will be a great daily driver model that you can use to get basic tasks done. — 10:29

We simply don’t have evidence to make that claim — 12:04

Concepts and entities

Gemini 4 Argon, Google DeepMind, Fairwind, GPT-6.1 Soul, GPT-6 Astra, Opus 5.5, Artificial Analysis Intelligence Index, AutomationBench, LV Bench, Reuters, ASI

Full transcript (original language)

00:00 Honestly, this space is so wild that if you are to even blink, you will miss a model update. Cuz today, Google has officially released Gemini 4 Argon, or Aragon if you are a huge fan of Tolkien’s work. But, this model is absolutely insane. One thing I want to mention right off the bat is that Argon is currently in a limited release to selected cybersecurity partners and trusted testers through Google DeepMind’s Fairwind program. But, Google is saying that they’re going to be having a broader rollout soon with paid

00:31 API customers and Google AI Ultra subscribers. So, hopefully we see it within the next couple of days, or maybe even next week. But, something I really want to talk about about this model right away is that this may be the best cost-to-performance model that we have ever seen. I know we just got GPT-6.1 Soul, but this model Argon is launching at an introductory price of just $2 per 1 million input tokens and $10 per 1 million output tokens with the cached input price 95% cheaper, which is

01:05 insane. And probably the craziest part is that Google has increased the maximum output from 64K to 1 million tokens output. That is an absolutely ridiculous amount of output for these massive long-running agentic and coding tasks, which is also supplying a low hallucination rate, which is going to make this model a great daily driver. Now, sorry to interrupt, guys, but this would truly mean a lot to me if you guys can go ahead and like this video and subscribe as it really helps me make videos for you guys on a day-to-day

01:36 basis. Let’s take a look at the benchmarks. Gemini 4 Argon may not be the best agentic coding model, but it may be the best text arena model that is available right now. This is something that surpasses GPT-6 Astra, Fable 5.1, and even Opus 5.5 within that domain on knowledge work. It also has that extended output limit with 1 million tokens, which is going to allow it to potentially reason and work through hundreds of thousands of tokens in a single trajectory. Now, one ridiculous

02:08 benchmark that I found about this model is Deep Sway 1.1, which is where it scores a 77.9%. It sets a new state-of-the-art record for real-world long-horizon software engineering tasks. It also is leading in Balz Index covering real-world work across finance, coding, legal, and tax. On Automation Bench, this is where it is able to complete end-to-end business workflows quite well, and Argon actually ranks number one with 51.3%.

02:40 And it even does quite well for long video understanding, where it sets a state-of-the-art record of 91.7% on LV Bench. So, it is setting these crazy records. It even does quite well on stuff like computer use with its efficiency, but I still find GPT-6.1 Soul better than this new model in that specific domain. But when it comes to multimodal understanding, this is where it surpasses any other state-of-the-art model. If you want the best AI tools, workflows, and drops before everyone

03:11 else, join my free newsletter with the link in the description below, which is completely free. Gemini 4 Argon also looks really competitive in terms of overall cost efficiency, where it costs just about $1.99 per Intelligence Index task. Compared to GPT-6.1 Soul, which has the cheapest pricing in this domain listed at 72 cents, GPT-6 Astra at $3.26, and Opus 5.5 at $5.98, and then Fable 5.1 at $7.63.

03:43 And Argon scores a 53 on the Artificial Analysis Intelligence Index, exactly matching GPT-6 Astra and sitting one point behind GPT-6.1 Soul. What’s crazy is that Argon achieves that while costing just 60% as per task as Astra as its introductory pricing though. Even when the promotion ends, pricing moves from the introductory pricing that we listed before to $4 per 1 million tokens and $20 per 1 million output tokens, which is still extremely competitive for

04:14 a frontier model. The gentic performance is also a massive improvement. Argon ranks number one on automation bench at 77.5% beating Sonnet 5.5 at 71.3% while also making huge jumps on turnal bench. And probably one of the most impressive results is hallucination because Argon has a 15% hallucination rate, the lowest on artificial analysis has ever measured for any model scoring 45 plus on the intelligence index. For

04:45 comparison, Astra is at 51% and GPT-6.1 Soul is at 54%. So, Argon isn’t absolutely the cheapest frontier model, nor is it the best, but when you combine the intelligence, the gentic performance, the low hallucination rate, obviously the token output, pricing, price to performance, everything looks seriously competitive and it is something that may even take the ranking for me as my daily driver. Another interesting detail that came

05:16 from Reuters is that Gemini for Argon is reportedly larger than Google’s previous Gemini for models suggesting that Google has scaled up the model itself alongside the massive jump in reasoning and the gentic capabilities, which is kind of insane. Perhaps this is a 6 trillion parameter model, maybe even more. Not too sure yet, but that is obviously an impressive feat. Now, like I had mentioned, this is a model that is currently only available to a select cohort of cyber defenders and trusted testers and they are going to be

05:48 allowing users like paid API users within the next couple of days, maybe even weeks, to get access to this model first, and even Google AI Ultra subscribers, and then work on a rollout to the broader public, maybe mid-October. I’m not too sure yet. I am just guessing. I believe maybe even in a week we could see this model. And luckily for you guys, I got someone from the AI Leaks official Discord sending me a couple of tests I’m going to be showcasing throughout today’s video. The model had actually generated

06:19 this entire Sakura Hokota Sanctuary using over 106K voxels, and the level of detail is genuinely incredible. From the atmosphere to the environment to the different camera angles, everything looks insanely polished with this output. Now, you can even dynamically change the time of day like any other voxelar, and it completely transforms the lighting and the atmosphere of the scene. Here you can see an SVG of Gemini 4, and

06:50 this honestly looks really good in terms of the output. What it was requested to create appears PS5 controller in SVG, and it actually looks quite real. It really looks like a real image to me, other than the texture and that small little gradient that you see on the left and right side of the ends of the controller. But other than that, everything looks quite well and accurate to the real controller itself. The lighting, the hue, everything has been generated quite well, and honestly it is comparable to the results of Opus 5 with

07:22 the PS5 controller. Here is another crazy comparison of the Gemini 4 Pro within the arena. This was an earlier checkpoint. Not sure if this is going to be Argon exactly, but this is a checkpoint that was put up against Claude Opus 5.5 tasked to build out a 3D float plane with physics simulation. This is where the Gemini 4 had completely outperformed Opus Swap on 5 within this test with the visuals, overall physics, and the execution, which is obviously on another level, and

07:53 you can tell from this generation. And if this representation of the model is something that they end up releasing with Argon, maybe even improving it over the next couple of weeks, that would be a great additional benefit, especially with that pricing. Here is another preview of this checkpoint where it is something that had generated this 3GS simulation of different characters for Halloween, and this is where you can see the level of detail being outputted from this model, and it truly looks next level when you

08:24 are comparing it with other models. This is honestly amazing cuz this is something that adds in another level of detail and depth with these simulations, and it’s just truly incredible. Here is where Gemini 4 was requested to create a mechanical bee, and this is honestly really incredible. This is something that has a really nice design taste correlated with this model. It does quite well with front end, which is something that the earlier Gemini models used to do quite well at, but then

08:55 Claude had overtaken that category, but now we can see that Gemini’s making that comeback, and this is a level of detail that I truly didn’t expect from the Gemini 4. This is obviously due to the prompt adherence that was attributed to this generation, but the level of quality that it was able to follow through with is awesome. Here Gemini 4 had created the Colosseum, and this is something that is honestly pretty remarkable cuz you also have different day and night cycles being simulated

09:25 with this generation. You can see the design quality and the polish here has truly improved in terms of the pro models that we see with Gemini, and it just feels significantly more refined with its overall output. Obviously, this is the previous Gemini checkpoints that we had saw, but still if it is a step towards Argon, it honestly shows us that Argon will be a significant leap and obviously a nice comeback from Google cuz these are generations that we

09:57 truly didn’t expect from Gemini after that small little stint being behind. Something that I have also realized with this model is that it is a lot better when it comes to interactive demos. It is more intuitive with these outputs and it even uses learning methods that we haven’t really seen from other models being outputted with these generations. It is also quite fast, but in terms of efficiency, let’s see if they end up sticking with this introductory pricing cuz that would be a huge steal and that

10:29 would essentially make everyone want to use Gemini models once again due to obviously the pricing performance. But I will say the Gemini for Argon will not obviously be the best coding or web development model available, but it will be a great daily driver model that you can use to get basic tasks done. And obviously, one area that Gemini usually excels at is multimodal capabilities. That is something that you can use this model for as your daily driver. It is not going to be able to compete with GPT

11:01 as well as Sonnet or even Opus due to those models already perfecting a lot of the capabilities within coding as well as in other domains. But hopefully, we see a recovery from Gemini and that they can excel in that category in the near future. And to be honest, this checkpoint does really correlate to what we see with Argon in terms of the benchmark scores cuz it is putting up a great fight with models like GPT-6 Astra. Here is a side-by-side comparison using the exact same prompt against

11:33 GPT-6, and you can see that Gemini 4 looks absolutely insane with its 3D world building using 3js to fully output this sort of simulation based off of what it was requested to build, like the train tracks, the dynamic world, even using different sorts of colors and textures to get this output. If you like this video and would love to support the channel, you can consider donating to my channel through the Super Thanks option below. Or you can consider joining our private Discord where you can access

12:04 multiple subscriptions to different AI tools for free on a monthly basis, plus daily AI news and exclusive content, plus a lot more. Overall, Gemini 4 Argon definitely looks like a massive release for Google and the AI community, obviously. And at the pace where Gemini is shipping all of these models, it is definitely getting ridiculous, and it honestly leads you to think if they have achieved ASI. We simply don’t have evidence to make that claim, but when AI systems are increasingly being to use to

12:34 internally help research and optimize, as well as code, evaluate, and ultimately build the next generation of AI systems, like they have claimed, you can start to see that the feedback loop could accelerate the progress of newer models. Google itself says that Argon is already helping its researchers and complex engineers with algorithmic work, as well as other areas within the company. So, that is obviously increasingly allowing you to believe that that claim is accurate. But

13:05 overall, I really love what they have done with Argon, and this is honestly a great step for Google and obviously making a nice step to making a comeback for the company in the AI race. So, really love what they have done with this model. Hopefully, they can keep on improving it so that it is something that can excel in almost every domain, like the state-of-the-art models that we have right now. But, it’s a huge upgrade within text, as well as in domains like multimodality. But, regardless, I am really hyped to see what this model ends

13:37 up being, and we will be making a full test video once it ends up releasing. I’ll leave all the links that I use in today’s video in the description below. Make sure you go ahead and take a look at the World of AI School community. Take a look at the World of AI Vibe coding platform and benchmark. Take a look at our second channel. Join us on on on on our newsletter. Make sure you go ahead and join the Discord. Follow me on Twitter. And lastly, make sure you guys subscribe, turn on the notification bell, like this video, and please take a look at our previous videos so that you can stay up to date with the latest AI news. So, with that thought, guys, have an amazing day. Spread positivity, and

14:08 I’ll see you guys really shortly. Peace out, fellas.