同一动画 prompt 五模型横评：M3.1 Flash 52 秒免费完成

AI Penguin | AI Routing @HeyAIpenguin原推发布于 2026.10.20

在 X 上打开原推

The US just ordered every civilian federal agency to rip out DeepSeek. So I gave it and 4 other AI models the same prompt. 🚀 DeepSeek: 17 seconds, $0.004. Claude Opus 5.5: 4 minutes, $0.50. That's 131× the price. The idea: Everyone is arguing about DeepSeek today: the DHS order, China's response, the "it's just as good" takes. Benchmarks are abstract. I wanted to see it on one real task, side by side. What I did: 1. Wrote one prompt: an animated race between a SpaceX rocket and an Anthropic rocket, $150 vs $660 per American kid, as a single HTML file 2. Sent it to 5 models through OpenRouter, once each, no retries: DeepSeek V4.1 Flash, MiniMax M3.1 Flash, Qwen 3.8 Max Prime, Gemini 4 Argon, Claude Opus 5.5 3. Recorded every result exactly as written. No fixes to the code 4. Costs are OpenRouter's own usage numbers, not estimates Results: DeepSeek V4.1 Flash · 17 s · $0.004 Works first try. Blocky rockets, and stray shapes left on screen at the end MiniMax M3.1 Flash · 52 s · free Whole story is there, but tiny rockets and an almost empty first second Qwen 3.8 Max Prime · 66 s · $0.09 Solid motion, but labels crash into each other and the title is cut off Gemini 4 Argon · 173 s · $0.14 Clean and polished. Simpler rockets, no cash piles Claude Opus 5.5 · 247 s · $0.50 Best by a mile: cash piles, real flames, a finale over Earth. One overlapping number at the end What didn't work: → My first DeepSeek run burned tokens on "thinking" and returned nothing. Turning reasoning off fixed it → Not one model shipped a flawless frame. Every one needs a human pass What I learned: 1. All 5 models produced a working animation on the first try 2. The cheap Chinese models are drafts. Opus is the final cut 3. Gemini gives you ~80% of Opus for under a third of the price 4. Best workflow: iterate on DeepSeek for $0.004, finish with Opus Which one would you ship? 👇

这条是引用着另一条写的：x.com/polymarket/status/2112528073828520016