Models

† Video is decoded and sampled at the model's native rate of 2 frames per second across the entire clip — up to 1,370 frames, i.e. clips as long as 685 s — with every sampled frame tokenized by the vision encoder (~64 tokens per second of 360p video). Prompt tokens therefore grow linearly with clip duration: a 10-minute clip is ~38,000 tokens of visual signal. Many endpoints instead cap videos at a fixed frame budget (~30 frames total, one frame every ~20 s on a long clip) regardless of duration. Context is 262,144 tokens, so even the longest clip fits comfortably. ZDR: zero data retention — prompts and media are processed in memory and never persisted or trained on; billing keeps token counts only. Each model page shows measured throughput and TTFT by clip length.