A robot does not need an essay about the scene. It needs to know where the cup is. Perceptron Mk1.5 is built around that difference. Released on 25 September 2026 as the successor to Mk1 from May, it is an embodied reasoning model for physical agents, and its answers can carry geometry: points, boxes, polygons, video clips and object tracks, alongside ordinary text.
Perceptron says the model already runs on drones, robot dogs, smart glasses and phones.
New since Mk1
- Native audio input, so a device can hear as well as see
- Video object tracking, following an object across frames
- Web search and sub-agent calls, letting the model fetch facts or delegate work
Tool definitions use the OpenAI format with JSON Schema, which means most agent frameworks can drive it without an adapter. Through OpenRouter, a reasoning_effort control is exposed.
Specs
| Item |
Detail |
| API ID |
perceptron-mk1.5 |
| Inputs |
Text, images, video, audio |
| Outputs |
Text plus optional points, boxes, polygons, clips, tracks |
| Context |
32K (launch post) or 36,864 (listings) |
| Max output |
8,192 tokens |
| Price |
$0.15 input / $1.50 output per 1M tokens |
| Reasoning / tools |
Yes / yes |
Annotations are billed as ordinary text tokens. A dense set of polygons can be long, so watch the output count if you request every object in a busy frame.
Two numbers to reconcile
The launch post says 32K tokens of multimodal context; listings say 36,864 with 8,192 output. A reasonable reading is that 36,864 includes the output budget, though we cannot confirm that. Design for 32K of input. The same post gives inconsistent speed-up figures over Mk1, from 2 to 5 times, up to 4.7 times, and 5 times, so do not plan around any single one.
A publisher-reported result
On egocentric hand localisation, Perceptron reports 0.9433 against 0.6179 for the best Gemini model it tested. First-person hand tracking is a core skill for glasses and manipulation robots, which is why Perceptron leads with it. It is a vendor figure.
Prices above are indicative; check your account. On-premises deployment is available through Perceptron's sales team.
Trying it on your robot's footage
Collect twenty clips from your own device's camera, at its real resolution and frame rate, with the lighting it actually sees. Ask grounded questions ("box every tool on the bench", "track the person in the red jacket") and overlay the returned geometry on the frames. Measure how far boxes drift over a clip, not just whether the first frame is right. Compare with Gemini Robotics-ER 1.6 Preview, and see other video-text-to-text models.