The Bellevue startup says its new vision-language model can process factory video at a fraction of the cost charged by OpenAI and Google
A startup barely two years old wants to give factories the ability to see. Perceptron AI, founded in late 2024 and headquartered in Bellevue, Washington, launched its flagship Perceptron Mk1 vision-language model on May 12, positioning it as a purpose-built AI system for industrial environments like manufacturing floors, warehouses, and construction sites.
The pitch is straightforward: take the kind of visual intelligence that powers chatbots and self-driving cars, strip away the consumer frills, and point it at the unglamorous but enormous problem of making physical workplaces smarter. Think defect detection on assembly lines, safety gear compliance checks, inventory tracking, and robotic navigation, all powered by a model that processes video natively rather than treating it as a series of disconnected still images.
What the Mk1 actually does #
The Perceptron Mk1 is designed around native video processing, which means it can handle temporal reasoning (understanding how things change over time), spatial understanding (knowing where objects are relative to each other), and embodied reasoning (figuring out how physical agents should move through a space). Those capabilities matter in a factory setting where a static snapshot tells you far less than a few seconds of footage.
The model processes video at up to 2 frames per second and operates with a 32K-token context window. For the non-technical: that context window is essentially the model’s short-term memory, determining how much visual and textual information it can hold in mind at once while making decisions.
Perceptron AI claims the Mk1 matches or exceeds frontier-level benchmarks in image and spatial reasoning when compared against leading models from Google, Anthropic, OpenAI, and Qwen. CEO Armen Aghajanyan and co-founder Akshat Shrivastava have framed their mission around making the physical world legible to AI systems.
The model ships with a Python SDK that supports detection capabilities and structured outputs tailored for industrial applications. An additional Egocentric API designed for hand-tracking is expected to launch in July 2026, which could expand the Mk1’s utility into areas like worker safety monitoring and manual assembly verification.
The price war angle #
Perhaps the most attention-grabbing detail isn’t what the Mk1 can do but what it costs. Perceptron AI says its pricing runs 80-90% below comparable models from the major AI labs. The numbers: $0.15 per million input tokens and $1.50 per million output tokens.
To put that in context, running visual AI at industrial scale has traditionally been expensive enough to limit adoption to the largest manufacturers with the deepest pockets. A small or mid-sized factory running continuous quality inspection across multiple camera feeds can burn through millions of tokens daily. At Perceptron’s pricing, the economics shift dramatically.
What to watch #
The July 2026 launch of the Egocentric API will be an important milestone. Hand-tracking in industrial settings could unlock use cases in worker training, ergonomic assessment, and manual quality control that current systems handle poorly.
The 2 FPS processing rate, while adequate for many inspection tasks, may prove limiting for applications requiring faster visual throughput. And the benchmark claims against Google, Anthropic, and OpenAI will face real scrutiny as early customers put the Mk1 through its paces in messy, unpredictable real-world environments where lighting changes, cameras get dirty, and products don’t always look like they did in training data.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our