Perceptron Inc., a two-year-old AI startup, announced the release of its flagship video analysis model Mk1, pricing it at approximately 80-90% less than competing models from major tech companies. The company charges $0.15 per million input tokens and $1.50 per million output tokens through its API, compared to significantly higher costs for models from Anthropic, OpenAI, and Google.

The Mk1 model is designed to analyze and understand video content, including live feeds, with applications ranging from security monitoring to content creation and quality control in manufacturing. According to Perceptron, the model can process video at up to 2 frames per second across a 32,000 token context window and maintains object identity even through visual obstructions.

Perceptron reported benchmark results showing Mk1 achieved competitive or superior performance across several industry-standard tests. On spatial reasoning benchmarks, the model scored 85.1 on EmbSpatialBench, surpassing Google's Robotics model at 78.4. In video analysis tests, Mk1 scored 88.5 on VSI-Bench, which the company claims is the highest recorded score among compared models.

The startup was founded by Armen Aghajanyan and Akshat Shrivastava, both former researchers at Meta's AI research lab. The company, based in Bellevue, Washington, focuses on what it calls "physical AI" - models designed to understand real-world video and sensory data for applications in robotics, manufacturing, and security.

Perceptron operates a dual-track approach with its AI models. While Mk1 is available only through API access, the company also offers an open-weights alternative called the Isaac series, with Isaac 0.2 designed for edge deployments requiring sub-200 millisecond response times. Early applications include automated sports highlight creation, robotics training data curation, and manufacturing quality control systems.