Perceptron Launches Mk1.5, a Multimodal Perception Model for Physical Agents with Structured Spatial Outputs

Perceptron has released Mk1.5, an update to its perception model line built specifically for physical agents — systems that need to understand and act in the real world rather than just process text.

The model accepts four input modalities: text, image, video, and audio. It responds with natural-language text and, when requested, structured spatial annotations: points, bounding boxes, polygons, and temporal clips. According to Perceptron, these annotations are only emitted inline with text when a caller explicitly sets the annotation_format parameter to "point," "box," or "polygon" for spatial localization on images, or "clip" for temporal segmentation in video. Without that parameter, the model returns plain text only.

Mk1.5 also supports graded reasoning through standard reasoning controls, letting developers trade latency for deeper analysis on harder queries. It supports function tool calling and structured outputs via JSON Schema. Audio analysis on video soundtracks is opt-in and only runs when explicitly enabled per request.

Pricing and specs

• Input: $0.15 per 1M tokens
• Output: $1.50 per 1M tokens
• Context window: 37,000 tokens
• Released: September 25, 2026 (per OpenRouter listing)

On OpenRouter, the model is currently showing 100% uptime and availability over a trailing three-day window, with a median latency of 0.57 seconds and throughput of 93 tokens per second — the best figures among the providers OpenRouter tracks for this model, according to the platform's own routing data.

How it compares to Mk1

Mk1.5 succeeds Perceptron Mk1, the company's prior flagship vision-language model for video and embodied reasoning. Mk1 accepted image and video inputs paired with natural-language queries and specialized in video QA, summarization, event detection, point-by-example grounding, OCR on noisy real-world documents, open-vocabulary object detection and counting, and hand pose estimation. Mk1 shipped with a 33K-token context window at the same $0.15 / $1.50 per 1M token pricing.

Mk1.5 extends that foundation by adding audio as an input modality, expanding the context window from 33K to 37K tokens, and formalizing the annotation system (points, boxes, polygons, clips) as an explicit, opt-in output format governed by a single parameter rather than model-specific behavior. No benchmark scores for Mk1.5 have been published as of this writing.

What this means

Perceptron is positioning itself in a narrow but consequential niche: perception models for robots and other physical agents, rather than general-purpose chatbots. The emphasis on structured, machine-parseable outputs — points, boxes, polygons, timestamped clips — signals a model designed to feed directly into robotic control loops or automation pipelines rather than human-facing chat interfaces. The addition of audio input and opt-in soundtrack analysis in Mk1.5 suggests Perceptron is building toward agents that need to fuse sound with vision, such as detecting an alarm or verbal command alongside a visual scene. The 37K context window is modest compared to frontier text models, reflecting a design trade-off toward low-latency, high-throughput perception tasks rather than long-document reasoning. At $0.15/$1.50 per 1M tokens, pricing sits well below flagship multimodal models from larger labs, which may matter for agent deployments that call vision-language inference at high frequency.