Perceptron Mk1.5

Perceptron Mk1.5 is a vision-and-audio model from Perceptron, released on 25 September 2026 with a blog post and API release notes of the same date; OpenRouter listed it that afternoon and HuggingNews reported it the same day. It is built to act as the perception and decision layer of physical agents — drones, legged robots, smart glasses and phones — without retraining for each device. It accepts text, images, video and audio and answers with text plus structured annotations: points, boxes, polygons, time-stamped clips and object tracks that follow one item through a video. Compared with the earlier Mk1 it adds audio input (up to about 21.8 minutes per item), video tracking, function calling with any JSON-Schema tool, adjustable reasoning effort and structured output. The model card gives a 36,864-token context window shared by all inputs and 8,192 output tokens; the blog rounds this to 32K. The price on the model card is $0.15 per million input tokens, $0.0375 cached and $1.50 per million output tokens, the same figures OpenRouter shows. Perceptron reports two to five times lower end-to-end latency than Mk1 on one H100, and on its own egocentric tests a hand-localisation score of 0.94 against 0.62 for the best Gemini model it measured, while trailing Gemini 3.8 Flash slightly on EgoSchema (80.4 against 81.2 per cent). Those are Perceptron’s own measurements and are not stored as scores; no independent evaluation had been published when this was read on 26 September 2026. The weights are not public; Perceptron offers them under a commercial licence on request, and the API retired the older Isaac 0.1 and 0.2 models on the same day, so anyone still calling those needs to move.