GPT-6.1 Sol vs Opus 5.5. Half the standard input/output token price. I built a paper-train adventure: broken tracks, folding landscapes, and a seed that brings a city to life. Two routes. One mission. Both versions rendered with Minimax.

This is such a stupid and misleading comparison

this makes no sense. its just minimax. with different prompts?

Same tool, but I didn't write those prompts. Each model operated Minimax itself from one instruction. What you're comparing is how each one directs and iterates on its own.

which is just a prompt. you dont handle minimax. You feed it a prompt. you dont even need a fancy model for that . some cheap local model can create the prompt for you just fine and youll have the same output. Its still just minimax generating the video based on a prompt t2v

Is Opus 5.5 really the bottom one here? Bottom output has quite a few issues and I honestly prefer the top one.

The pricing on the new Sol model is actually quite a game changer for scaling:

This is a Minimax comparison, NOT an Opus vs Sol comparison 🙄

Fair point, but Minimax isn't the variable. I gave each model one prompt + a reference video and let it drive Minimax via computer use, start to finish. No edits from me. This is an agent test, not a renderer test. The flaws are part of the result.

You do realize that, unless you use the same exact seed, that every render is going to be different with Minimax, right? So Minimax is absolutely still a variable.

Randomness isn't a variable, it's noise, and it's the same for both sides. By that logic no LLM could ever be benchmarked.

看着6.1sol不错，不知道实际执行额度消耗如何

Claude fan boys are hurted

GPT-6.1 🚀

been seeing that gpt 6.1 is far better than opus 5.5 but this is almost similar

Probably because this tests computer use, not raw reasoning. Once the model is clicking through a real UI, the gap shrinks. Both got to a watchable video from one prompt.

Idk what to believe anymore

How do I create this animation?

One prompt + a reference video, then I let the model use Minimax on its own via computer use. That's it, no manual editing.

How to generate such videos

saw cheaper models cost more after retries, token price alone hides workflow cost

Agree, and it's even worse with computer use. Every misclick or re-render burns tokens. The metric I care about now is cost per finished video, not per token. Have you tracked that for agent runs?

这种任务大概只能测试模型的提示词写作能力，而且有很大随机性，模型并不知道什么提示词能达到好结果，官方的提示词写作指南显然并不够