AI视频提示词对比评测

GPT-6 Astra 刚刚发布， 它在制作元广告方面简直疯狂 但 Fable 5.1 仍然有一件事做得更好。 我们把两者都放进了同一个测试框架。相同的简报、相同的产品、相同的渲染器——Seedance 2.5。每个模型编写提示词。我们渲染两者。我们进行比较。 Astra 在正文部分胜出。差距不小。 更干净的场景过渡。产品在每个镜头中保持一致。节奏全程保持。Astra 广告的最后 7 秒看起来像是代理商真正交付的作品。 然后你看看前 3 秒。 Fable 的钩子能阻止滑动。Astra 的钩子只是开始广告。 我们对此的最佳解读是： Astra 是为完成工作而构建的。OpenAI 自己的发布文章说，它是他们最擅长遵循模板和产出结构化叙事的模型。精致。详尽。均衡。 这正是正文所需要的。 这与钩子所需要的恰恰相反。钩子故意是未完成的。它是一个问题，而不是答案。详尽会扼杀它。 给 Astra 一个 4 拍简报，它会写出 4 个相等的拍子。第 1 拍必须格外响亮。但它不是。 好消息：这可以在提示词中修复。 我们为 Astra 编写了一份提示指南，修补了它自带的默认设置。钩子优先结构。不对称规则。适用于 Seedance 2.5 制作元广告的确切简报格式。 评论“hook”，我会发给你。 两者都运行在 HeyOz 上——每个模型，并排，相同产品。

GPT-6 Astra just dropped, and it's INSANE for making meta ads BUT Fable 5.1 still does 1 thing better. We ran both through the same harness. Same brief, same product, same renderer — Seedance 2.5. Each model writes the prompt. We render both. We compare. Astra won the body. Not close. Cleaner scene transitions. Product stays consistent shot to shot. Pacing holds all the way through. The last 7 seconds of an Astra ad look like something an agency actually shipped. Then you watch the first 3. Fable's hooks stop the scroll. Astra's hooks start the ad. Our best read on why: Astra is built to complete work. OpenAI's own launch post says it's their best model at following templates and producing structured narratives. Polished. Thorough. Even. That's exactly what a body needs. It's the opposite of what a hook needs. A hook is unfinished on purpose. It's a question, not an answer. Thoroughness kills it. Give Astra a 4-beat brief and it writes 4 equal beats. Beat 1 has to be disproportionately loud. It isn't. Good news: it's fixable in the prompt. We wrote a prompting guide for Astra that patches the defaults it ships with. Hook-first structure. Asymmetry rules. The exact brief format that works with Seedance 2.5 for Meta ads. Comment "hook" and I'll send it. Both ran on HeyOz — every model, side by side, same product.