Perceptron releases Mk1.5 with full multimodality for physical agents at unchanged pricing

Perceptron has launched Mk1.5, a perception model built for physical agents, while holding the previous generation's pricing: 15 cents per million input tokens and $1.50 per million output tokens (roughly 0.55 shekels and 5.5 shekels respectively). The September 25, 2026 release arrives without a price increase, positioning the model as a capability upgrade rather than a generational leap that demands a new budget.

The core change is sensory bandwidth. Mk1.5 accepts text, images, video and audio, and returns text alongside optional structured annotations — points, boxes, polygons for spatial grounding in images, and clips for temporal segmentation in video. Annotations are not emitted by default; they appear inline with the text only when a developer explicitly requests them via the `annotation_format` parameter. A video's audio track is analyzed only when a dedicated per-request flag is enabled, a mechanism designed to avoid unnecessary compute and cost.

Beyond modality, the model supports graded reasoning through the platform's standard reasoning controls, function tool calling, and JSON Schema structured outputs. The combination lets a physical agent not just "see" but also plan steps, invoke external tools, and return output that interfaces directly with control systems — no client-side translation layer required.

The context window sits at 37 thousand tokens, modest by general-purpose LLM standards but sufficient for most robotics and computer-vision scenarios where visual input is compressed. Throughput measures 93 tokens per second (P50, best among providers), with a median latency of 0.57 seconds (P50, best provider). Those numbers put Mk1.5 in territory that enables near-real-time control loops, provided the request stays within the context limit.

The platform reports 100% availability thanks to automatic routing to a healthy provider when an upstream provider errors, as long as request filters permit it. Per-provider availability data is accessible programmatically via the Endpoints API, letting engineering teams build custom dashboards and define their own fallback policies.