Gemini 4 Argon - Out and Tested!!

Author: Sam Witteveen · 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 - Out and Tested!! · Channel: Sam Witteveen · 11:10 · Watch on YouTube

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

Summary

Sam Witteveen presents Gemini 4 Argon as a strong multimodal model, especially for tasks combining visuals, code, and media, while arguing that it is not necessarily the best model overall. Examples include generating a music-making interface and creating Blender animations through an iterative code-render-review loop. He says Argon is not primarily aimed at winning coding-agent comparisons or benchmarks; rather, it gives Google and its customers a capable frontier model within the Google ecosystem. He also explains that the million-token response capability is an API feature with specific configuration requirements, and concludes that Argon is promising but still evolving.

Key points

Argon excels at multimodal tasks that combine visuals, code, and media.
A short prompt in the Gemini app produced a music-making interface with synthesizers, drum machines, effects, and a step sequencer.
Blender and game-generation tasks appear to benefit from iterative workflows in which the model writes code, renders results, inspects them, and makes fixes.
Witteveen says Argon is not the strongest choice for agentic coding, and that other models may perform better on coding tasks.
He views Argon as a credible frontier model for Google customers and a foundation for Google products such as Gemini, NotebookLM, and Workspace.
The million-token capability is mainly for API use, requires a configuration flag, and is most useful for tasks needing hundreds of thousands of tokens of reasoning.
Witteveen recommends testing Argon on individual use cases rather than rebuilding systems around its long-context capability.
He says the model is still being trained and tested, and expects it may improve in a future checkpoint.

Chapters

Start	Chapter	Summary
00:00	What Argon is—and what matters beyond leaderboards	Witteveen introduces Argon as a very good model, explains his early access, and frames the video around its strengths rather than leaderboard position.
03:07	Multimodal strengths and iterative creation	He discusses Argon’s performance on Blender, game-generation, and other tasks that involve writing code, rendering, inspecting, and refining results.
04:43	How Argon compares and why Google built it	Witteveen weighs Argon against other models and argues that Google needed a capable frontier model for its customers and product ecosystem, not necessarily the top coding model.
08:26	The million-token feature in practice	He clarifies that million-token responses are an API-oriented option requiring configuration and are most relevant to unusually long reasoning tasks.
10:06	Conclusion, pricing, and future updates	Witteveen summarizes Argon’s multimodal value, mentions its introductory pricing and ongoing development, and invites viewers to share their views.

Quotes

This is a really, really, really good model. — 00:00

that’s a single-shot prompt. — 02:36

It’s that loop of sort of code, render, look, fix something up — 04:11

Argon delivers here — 06:50

Argon is a really good model. — 09:32

Concepts and entities

Gemini 4 Argon, Google, Antigravity, Gemini API, Gemini app, Blender, ReBirth by Propellerheads, Claude Code, Codex, Sam Witteveen

Full transcript (original language)

00:00 Okay, so Gemini 4 Argon is now out, And in this video, while I’m gonna show you a few demos, I’m gonna really talk about more about what this model means and how you should actually think about this model so I’m gonna say this right up front. This is a really, really, really good model. But is it the best model out there? I actually don’t think it is. But more importantly, I don’t think that that really matters that much, and I’m gonna talk about why in this video Because what you’re looking at here are some of the demos that I basically put together for this showing

00:31 where this model is genuinely strong. it’s not always gonna be the stuff that tops leaderboards And in some ways, I think it’s actually more interesting to understand why Google built this model. and that requires looking past leaderboards anyway so when Google first announced Argon, I made a short video about it then Sure enough, everyone was complaining that there weren’t many examples, the model wasn’t available, all that stuff. Totally fair comments, by the way. So full disclosure, I’ve been an early tester on this model, and I had to sign quite a few more NDAs than normal.

01:02 but by the time you’re watching this video, Gemini now is part of the paid tier of the Gemini API. it’s gonna be in Antigravity. It’s gonna be in the Gemini app. It’s gonna be in almost every Google surface where they’re putting models. Now often it’s only gonna be available if you’re paying for this model but for things like AI Mode and stuff like that, it’s not just available to Ultra subscribers, it’s also gonna be open to Pro But basically the cool thing here is I can finally talk about what it’s like to actually use this model.

01:32 Alright, so where is this model actually strong? for me, the standout things here is anything that combines sort of visuals, code, and media in one kind of task If you take a look at something like this, which I chose to do not in Antigravity, but just in the pure Gemini app, you can see here I’m basically saying, “Can you build me a music maker in the style of ReBirth by Propellerheads. Try to give it an 808, a 909, and a 303, and then put it together And make sure to include filters for FX and a Roland-style step sequencer.

02:05 Okay, I deliberately made the prompt here very short, right? To see, okay, what can this actually come up with? ? So I’ve got a reference to something like ReBirth. I’ve got a reference to some, Roland, synthesizers, drum machines in there. And sure enough, it went away and produced something which is really pretty amazing in here. if I come in and run this

02:36 Straight away you can see that It’s built not only a very sophisticated interface here, but it’s also then gone and actually made, the sounds to this and built a little synthesizer, built something like an 808 drum machine, a 909, et cetera and that’s a single-shot prompt. now before you jump down my throat about the real world is not about using just single-shot prompts, I agree with you, right? I just felt that some of these are actually really good to look at and get a sense of what the model can actually do with a limited set of resources.

03:07 And let me explain what I’m actually mean by that so the model is extremely good at doing things like Blender and doing things like game generation, those kinds of things. And clearly that’s come from a lot of the reinforcement learning environments that they’ve trained this model on And done a lot of work post-training it on and it’s not just one demo here. If you look at the demos that Google put out, with the launch themselves, There’s a pretty clear pattern. In one of them, Argon basically, scripts Blender animations, generates

03:39 custom audio, pipes the whole thing through to an actual vintage CRT TV. In another one of them, they built a multi-tiered, Japanese temple in the grounds using Blender. And if you look at that, that’s pretty amazing to see, what this thing can do when it’s given an agentic loop and a proper harness that can use those kinds of skills and all of this feels like it’s had a lot of RL on these kind of pipelines, being things like Blender, Three.js, shaders, et cetera And those are the kind of tasks where the model writes the code, renders something out, looks

04:11 at the result, then gradually fixes it. It’s that loop of sort of code, render, look, fix something up that continually seems to be doing well here now in this video, I don’t want to get into benchmarks too much But what I do wanna say is that from everything I’m hearing, This is the biggest model that they’ve done pre-training on. You’ll notice in the name being Argon, that it’s not actually falling into the other tiers that they’ve had before, whether that was, Flash, Pro, even Ultra, which sort of got discontinued And we do know for a fact that this is just the

04:43 first Gemini 4 model Now the big question people are gonna ask, Is it as good as the Anthropic models or the OpenAI models, or even things like the new Muse Wonder here? so I got to say in my testing, a lot of this now comes down to how you’re actually using the model. I still don’t think, for example, Antigravity is anywhere near as good as Claude Code or Codex And as of the time recording this, I haven’t tried this in AI Studio Build, for example. I’ve basically been either using it from the raw API in

05:13 Antigravity or in the Gemini app. but clearly the thing here is that It’s definitely good at a lot of these multimodal tasks. Now, In the last video, I did talk about some of the benchmarks from the Artificial Analysis Intelligence Index, where this model scores a 53, which ties it with GPT-6 Astra and Fable 5.1, but quite a bit behind Opus 5.5. and even the new Sonnet 5.5 added to that in the mix is now Meta has released Muse Wonder, which also is quite a bit ahead of this model but one of the things

05:45 I will say here is that the benchmarks don’t always tell the full story, right? This model seems to be very good at generalization. I will say it’s not the best coding model. I think the other models are probably gonna do better at different kinds of coding tasks and for the whole sort of coding agent game. This model though really isn’t aimed at taking those, models out at their own game. so I think if your main job is picking the strongest model for agentic coding, you’re probably gonna look elsewhere. and I’m sure a lot of AI engineers are gonna complain about things

06:16 like that, and that’s a fair call. But I don’t think that’s why Google built this model. I don’t think Google’s actually trying to be in the game of being a standalone company that just sells tokens if you look at the Google for Business event last week, they basically announced that they had over 500 companies now that had done more than a trillion tokens with the Gemini models. my bet is most of those are actually the Flash models, perhaps even the Flash-Lite models what Google was lacking was a frontier model that they could basically put out there to their customers and say, “We also have a very good model.”

06:50 and that’s clearly what Argon delivers here it allows their customers to basically have a credible option of using this model if they’re already a Google shop or they’re looking to use Google products now Gemini 4 Argon gives those people a credible option without having to leave the Google stack And I think in many ways that’s the real customer here and then on the consumer product side, Google’s needed a frontier model underneath everything that it ships. The Gemini app now has tasks that keep running when you’re away from your computer.

07:21 It can connect to all these different apps. You’ve got NotebookLM, you’ve got things like Antigravity, And all those things are gonna benefit not only from this model, but probably from distillations of this model to make other Gemini 4 models having Argon allows the product people at Google to really focus on building the products that are working, things like NotebookLM, and probably improving the ones that perhaps haven’t gotten as much traction as they would like for things like Antigravity And putting this in the Gemini app, the Google Workspace apps, et cetera so I think in many ways,

07:55 basically Google here is focused on Argon as being a foundation and in that sense, this is clearly a big win for them while this year they’ve lost a lot of the top talent that came up with a lot of the ideas for a lot of the early LLMs Argon really shows that they’ve still got people there that know how to use the TPUs, know how to spin up a lot of them, know how to do a lot of pre-training, know how to do the RL environments and know how to put it together to actually make a frontier-level model so I think the best thing is rather than me show

08:26 you a lot of demos about what it can do, what it can’t do, Really at this point, you should be taking your use cases into either AI Studio or the Gemini app, or if you want to, Antigravity, and actually running those things yourself. I will say that in the first video, I talked a lot about the million tokens, in a single response. In practice, it turns out it’s a lot more nuanced than I originally had thought. the key thing there is that’s not gonna be in the Gemini app. I don’t know if it’s gonna be in Antigravity. Really, it’s more for the API, and it’s actually a flag that you need

09:00 to pass in, to the config, not to actually get it to keep going. and really you’re only going to benefit properly from it if you’ve got a task that requires multiple hundreds of thousands of tokens for reasoning and while I think that’s gonna be interesting to see how people prompt it to get it to just keep reasoning and reasoning and reasoning and go into that sort of loop in its reasoning I don’t see that that’s gonna get a lot of mainstream use So I’d still call this sort of early days. if you’re building your own vertical harness and stuff like that, I probably

09:32 wouldn’t go out and redesign everything around the million tokens yet that said, if you do have scientific use cases or things like that where you can actually get really long reasoning out of it, This is certainly worth turning on for your long reasoning tasks and measuring it to see actually how it plays out for you. anyway, to wrap up, Argon is a really good model. it may not be at the top of the leaderboards, and I’m sure many people are gonna criticize it for that But I think the real win for Google here is just what they themselves are able to build on it, and also their customers are able to build on it for

10:06 different use cases and the Google models have always been very good at generalization for these kind of tasks. they haven’t gone out trying to make the best coding model or to win every single benchmark that they’re going after. What they’ve made is a very, very strong multimodal model that has a really good sense of a lot of intelligence that perhaps was more on the fringe in the previous models Anyway, don’t forget that the price currently is $2 in and $10 out. That’s gonna last till December 1. And honestly, I’m gonna say I don’t think you’re gonna see this

10:38 model still around by December 1. We know that this is not the final checkpoint of this model. Still doing a lot of training , doing a lot of tests to how to improve it just like you saw with GPT-6 Sol, getting a massive bump when it went from 6 to 6.1, my guess is that Argon is also gonna get that same kind of bump in the next month or two. so anyway, I’ve put the links to the Google blog post and the Artificial Analysis in the description. if you missed my first video on Argon, check that out. let me know in the comments what you think and whether you think it really matters

11:10 that this is not the number one model And as always, if you found the video useful, please click like and subscribe, and I will talk to you in the next video. Bye for now