4 模型各给 $5 建 20 个网站：M3.1 Flash 全建成，其余三家破产
同一 agent 同一 prompt，M3.1 Flash 用 $5 建完 20 个手机可用网站还剩 $1.90，Gemini 3.8 Flash 建 2 个、GPT-6 Luna 9 个、DeepSeek V4.1 Flash 13 个后耗尽预算；M3.1 Flash 还主动跳过已有网站的两家。
i gave 4 AI models $5 each and told them "build $30,000 of websites or you die" that's what a local agency quoted me for 20 sites. one model built all 20 and still had $1.90 left the other three went broke same coding agent, same prompt, same 20 businesses on google maps with 4.8 stars and no website. i only swapped the model. every API call came out of its own $5 → Gemini 3.8 Flash: dead after 2 sites → GPT-6 Luna: dead after 9. cheapest per token in the race, but half its sites broke on a phone, and every rebuild came out of the same $5 → DeepSeek V4.1 Flash: dead after 13 → MiniMax M3.1 Flash: all 20 sites, every one works on a phone the winner is @MiniMax_AI's new M3.1 Flash. it launched yesterday and a lot of you already used it without knowing: for 10 straight days it was the #1 model on OpenRouter under a fake name, Space Bunny 🐰 i also hid 2 businesses that already had a website. only M3.1 Flash checked before building and skipped them and it's not just my test. on DeepSWE, a public benchmark of real coding tasks, it scores the same as Gemini 3.8 Flash for about 1/9 of the cost per task if you want to run it yourself, it's unlimited in MiniMax Code (their coding app) until oct 25 next run: $50 and 200 businesses. how many do you think it finishes before it dies?