Model / API profile · Checked September 27, 2026
Perceptron Mk1.5
Perceptron Mk1.5 is a developer model/API that accepts text, images, video and audio and can return text, spatial annotations and tracking outputs. Its embodied-agent positioning does not make it a packaged consumer editor or proof of real-world robot performance.
Release: 2026-09-25. Source-verified; not independently benchmarked.
Capabilities and reader fit
The official documentation describes points, bounding boxes, polygons, clips and tracks alongside language output and function calling. This fits developers prototyping perception, video analysis and agents that must refer to objects or events. The application still needs to execute tools and interpret the returned geometry.
Current access and exact model
The public developer platform and SDK document perceptron-mk1.5 with a 32K context window. An official demo is also linked, so “API-only” would be too absolute. This profile concerns the API/model workflow; it does not claim an end-user editing suite or downloadable public weights.
API pricing
The documented rate is $0.15 per million input tokens and $1.50 per million output tokens, with cached input at $0.0375 per million tokens. This is usage-based API pricing, not a subscription. Budget for the actual media tokenization, output length and application workload; this profile does not translate those figures into an untested cost per video.
Limitations and evaluation
Spatial predictions and tool calls need application-side validation. Model output is not authorization to move equipment or a guarantee of safe robot control. Test an authorized short clip with known objects and occlusions, inspect boxes and tracks frame by frame, and log the tokens billed. We have not tested latency, tracking quality or robotics benchmarks. For infrastructure-specific hosting, contact the provider rather than assuming self-hosting is included.
Official sources
Checked September 27, 2026. Access and prices can change; follow the linked provider documentation before use.