Perceptron AI launched Mk1, its first proprietary video understanding model, on Tuesday. The Bellevue startup priced it at $0.15 per million input tokens and $1.50 per million output, which the company says undercuts Anthropic's Claude Sonnet 4.5, OpenAI's GPT-5, and Google's Gemini 3.1 Pro by 80 to 90 percent.
The model processes video at up to 2 frames per second across a 32K context window, per the company's launch page. It tracks objects across frames through partial occlusions and emits structured annotations: bounding boxes, polygons, and timestamps that downstream robotics systems can consume directly. Reasoning can be toggled on or off per request.
Perceptron's own benchmarks put Mk1 at 85.1 on EmbSpatialBench against 78.4 for Google's Robotics-ER 1.5, with similar margins on VSI-Bench and RefSpatialBench. The numbers are self-reported, and independent testing hasn't followed. The VentureBeat writeup notes the comparison chart excludes saturated general visual-language tests, which narrows the field where Mk1 looks dominant.
The company was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both former Meta FAIR researchers who co-authored the Chameleon and MoMa multimodal architecture papers. "We built Perceptron to make the physical world legible to AI systems," Aghajanyan said in a press release. That's the standard pitch. The harder claim is in the cost chart.
Target customers include manufacturing inspection, security cameras, warehouse inventory, and robotics teleoperation data labeling. Mk1 is available through the Perceptron API and on OpenRouter. The open-weights Isaac series, launched in 2025, continues as a separate track.
Bottom Line
Mk1 sits at $0.15 per million input tokens, roughly one-tenth of Gemini 3.1 Pro's blended cost on Perceptron's chart.
Quick Facts
- Pricing: $0.15 per million input tokens, $1.50 per million output
- Video processing: up to 2 FPS, 32K context window
- EmbSpatialBench score: 85.1 (company-reported)
- Founded November 2024 in Bellevue, Washington
- Founders: Armen Aghajanyan and Akshat Shrivastava, both ex-Meta FAIR
