Perceptron launches Multilook API for cheaper multi-question video queries

Perceptron AI, which builds models that answer questions about visual input, launched a Multilook API that, per its announcement, prefills a video into the model's context once per API call instead of once per question. Prefilling is the step where a model reads and processes its input before generating an answer; normally that read happens fresh for every separate question, even against the same video.

Perceptron reports that sending up to 16 questions about one video in a single Multilook call costs 32 percent of what separate calls would cost, with median latency 2x to 4.9x lower under load. Those are Perceptron's own reported figures from its release, not independently verified by outside benchmarking, but they describe a concrete architecture change, sharing one context across many questions, rather than a vague efficiency claim.