Astra 不吃视频输入，用抽帧加 JSON 绕过
实测发现 Astra 只认图像，于是按固定帧率抽帧、烧入时间戳、强制 JSON 事件输出，在体育与仓储场景跑通。
GPT-6 Astra doesn't take video input... turns out that barely matters it only sees images, so the workaround is to feed it frames instead, here's how it actually works: - pull frames at a fixed rate, burn the timestamp into each one - send them as a sequence of images in one request, not a video file - force strict JSON output with an event schema, not free text - reasoning effort high matters, accuracy drops without it - ran it on sports analytics, kitchen process checks, and loading dock turnaround - tennis test: 7 serves and 17 hits caught even at 1fps, point outcomes need 4fps minimum - cost stays low, roughly a dollar per hundred frames the real trick for production: don't run astra on everything, let a cheap detector model watch the feed and only call astra for the moment that actually needs reasoning