핵심 요약
MiniMax M3라는 가성비 끝판왕 AI 모델이 나와서 Opus 4.8, GPT 5.5랑 빡센 아케이드 게임 원샷 빌드 대결을 붙여봤다. 프론티어급 코딩, 100만 토큰 컨텍스트, 멀티모달을 다 지원하는데 가격은 거의 공짜 수준이다. 실제로 돌려보니까 지가 알아서 모바일 테스트까지 하고 작업 검증도 젤 빡세게 해서 결과가 아주 충격적이다.
주요 포인트
- MiniMax M3는 코딩 능력, 100만 토큰 컨텍스트, 멀티모달을 동시에 갖춘 최초의 가성비 모델임.
- 성능 검증을 위해 Claude Opus 4.8, GPT 5.5와 함께 복잡한 60fps 네온 웨이브 슈터 게임 원샷 빌드 테스트를 진행함.
- 요구한 게임 스펙은 외부 라이브러리 없는 단일 HTML 파일 구동, 파티클 효과, 사운드 자체 제작, 모바일 최적화 등임.
- MiniMax M3는 실행 비용이 거의 안 드는데도 지가 알아서 데스크톱이랑 모바일로 교차 검증까지 마치는 집요함을 보여줌.
MiniMax M3 just launched, and it might be the most important model release of the year for one reason that almost nobody is talking about. It is the first model to bring all three of the big things to the open world at once. Frontier level coding, a 1 million token context window, and fully multimodal, all in one model. And the insane part is that it costs almost nothing to run. So, I did the obvious thing. I took a pretty brutal and complex build, and I just gave it to the three best models out there right now. So, Opus 4.8, GPT 5.5, and MiniMax M3. So, it's the same prompt. It's word for word, and I tracked every single token. Now, the part that I did not expect is the model that tracked its own work the hardest, and that tested itself on desktop and on mobile before it called it done, and it is also the one that cost me basically nothing. Honestly, the only reason I even ran this is because so many of my school community members kept asking me about M3, since so many of them are using it. So, I reached out to the MiniMax team, and yeah, they are sponsoring this video, but the test is real, the prompt is real, and every number is going to be right there on the screen, so you can just be checking with me the whole way. Also, if you want to join my school community and join all my other members, then check out the link down below in the description. It's completely free. I have a bunch of different assets and a bunch of videos that I do not cover on my YouTube. Check that out. All right, so let me show you the test because this is pretty much the whole video. Now, one of the tests that I'm providing it is to be pretty hard, not just a simple to-do list, not another simple landing page. I wanted a real one-shot build with just a hundred different ways to be doing it poorly, so I asked for a game. So, what I'm saying is build a complete polished arcade game. I'm going to run this off, but just so you know what I actually asked for, it's just going to be a neon wave survival shooter. So, the whole thing in one HTML file, no libraries, no build steps. So, it has to just be run by just double-clicking it. So, it's going to be 60 frames a second on a canvas. Enemies that are just going to be coming in should be coming in waves and be chasing you. There should be particle effects, screen shakes, a combo multiplier, even including lives and sound effects built from scratch with code, with a high score that saves, and it has to work on a phone/mobile as well. Now, I would argue that that's a pretty complex and genuinely a hard ask. There's a ton of room to actually nail it or just sort of get there. So, as you would expect, we're running this with MiniMax M3 on high, but we're also going to be comparing this with Opus 4.8 and ChatGPT's 5.5. So, I'm just using this inside of Hermes. It's just easy way to navigate and use all these different models all at once. All right. Now, here's who's actually building this. So, you can see right here, I have Codex open. Now, on Codex, as you would assume, I am running this on GPT 5.5. In Claude, we are now running Opus 4.8. And over here, like I just showed you, we are running MiniMax M3 running inside of the Hermes desktop application. It's just the easiest way to be building this out. And I already use Hermes quite often. Now, real quick, I'm not going to be setting up Hermes in this video. I already made a whole video walking through the desktop application, how to install it, all of that stuff. So, it's going to be linked down below if you want to check that out. Here, we're just going to be going straight to the results. So, I used the same prompt across all three of these models. I let them cook for a little bit. And right away, you can also just feel them working differently side by side. So, GPT 5.5, from what I found, it just goes. It's really heads down and it's very clean. Opus, it writes a ton. It's over a thousand different lines. And M3, if you watch this, it just keeps on thinking. It keeps checking itself. So, I will say, just hold on to one question while we go through this entire thing. Why did the cheapest model end up checking its own work the hardest? So, again, just hold that thought. We're going to come back to that at the end. Okay. So, with these three models with the same prompt, here's what actually happened. I'm going to be starting with the most expensive one, Opus 4.8. And I'm just going to say it. The priciest model in this entire build is the one that gave me a black screen at first. So, I double-clicked the file that it built, and there was nothing. It was completely dead. But, and this is very important, watch what it did next because this is Opus being genuinely brilliant. So, I told it that the game does not work, and it figured out the exact problem on its own. So, when you open an HTML file, typically by just double-clicking it instead of running it on a server, the browser it just blocks it from saving the high score. And that one thing it