Perceptron Unveils Mk1, a High-Performance Video Analysis AI Model 80-90% Cheaper Than Anthropic, OpenAI, and Google

Perceptron Mk1 is a closed API model targeting video analysis and reasoning, priced at $0.15 per 1 million input tokens and $1.50 per 1 million output tokens. Perceptron emphasizes that it is approximately 80-90% cheaper than Anthropic's Claude Sonnet 4.5, OpenAI's GPT-5, and Google's Gemini 3.1 Pro, and stated that it spent 16 months redesigning a multimodal pipeline for handling the physical world.

Its benchmark results are also aggressive.

• It scored 85.1 on EmbSpatialBench, surpassing Google Robotics-ER 1.5 (78.4) and Alibaba Q3.5-27B (approximately 84.5).
• It scored 72.4 on RefSpatialBench, showing a large gap compared to GPT-5m (9.0) and Sonnet 4.5 (2.2).
• It recorded 41.4 on the EgoSchema Hard Subset and 88.5 on VSI-Bench, demonstrating temporal reasoning capability in video.
• On the efficiency frontier, it explained that performance is close to top-tier models while cost is around $0.30 blended cost.

Rather than the typical VLM approach of viewing segmented frames, the architecture is designed to process native video at up to 2 frames per second, maintaining temporal continuity within a 32K token context window. This allows it to preserve object identity even when occlusion occurs, and to return a structured time code for a specific moment in long videos.

The model also emphasizes Physical Reasoning, highlighting capabilities such as judging before/after a buzzer by viewing ball position and the shot clock together, reading analog gauges and clocks, and counting up to hundreds of objects in dense scenes. The Python-based Perceptron SDK includes Focus, Counting, and In-Context Learning, enabling tasks like PPE detection or crowded object counting to be handled via natural language instructions.

The strategy is two-pronged. Mk1 is a closed model provided via API, while the Isaac series remains open-weight. Isaac 0.2-2b-preview targets sub-200ms time-to-first-token, and founders Armen Aghajanyan and Akshat Shrivastava are from Meta FAIR. Perceptron is expanding this lineup into real-world use cases such as sports highlight clipping, robotics teleoperation data refinement, manufacturing quality inspection, and smart glasses assistance.