← Back to feed

Perceptron AI has launched Mk1, a video understanding and embodied reasoning model aimed at the gap between frontier capability and frontier pricing. The model, available via API and OpenRouter from May 12, benchmarks competitively with GPT-5 and Gemini 3.1 Pro on spatial and video tasks while priced 80-90% below either.

Benchmark Numbers

Perceptron published results across three benchmark classes:

Spatial reasoning (ER Benchmarks)

• EmbSpatialBench: 85.1 — above Google Robotics-ER 1.5 (78.4) and Alibaba Qwen3.5-27B (~84.5)
• RefSpatialBench: 72.4 — against GPT-5m (9.0) and Claude Sonnet 4.5 (2.2)

Video reasoning

• EgoSchema Hard Subset: 41.4 — matches Qwen3.5-27B, significantly above Gemini 3.1 Flash-Lite (25.0)
• VSI-Bench: 88.5 — highest recorded score among compared models

The RefSpatialBench numbers are the sharpest signal. General-purpose frontier models hit single digits on referring expression comprehension for spatial tasks; Mk1 scores 72.4. That gap reflects purpose-built training rather than incremental improvement on a shared base.

Pricing and Positioning

Mk1 is priced at $0.15/M input and $1.50/M output. Blended at 25% input / 75% output: $0.30/M. Perceptron compares against GPT-5 at approximately $2.00 blended and Gemini 3.1 Pro at approximately $3.00. The “Efficiency Frontier” chart Perceptron published plots Mk1 at roughly 10x lower cost than the nearest frontier competitor at equivalent benchmark scores.

Context window is 33K tokens. The model accepts image and video inputs with structured spatial annotations on demand — point, bounding box, or polygon — returned only when requested via the annotation_format parameter.

Who Built It

Perceptron was founded in November 2024 by Armen Aghajanyan (formerly Meta FAIR and Microsoft Research) and Akshat Shrivastava (also formerly FAIR). The company is headquartered in Bellevue, Washington. The launch post cites 16 months of development on a “multi-modal recipe built from the ground up.”

Alongside the closed-source Mk1, Perceptron maintains an open-weight Isaac series for edge deployment. Isaac 0.2-2b-preview, released December 2025, is a 2B-parameter vision-language model with reasoning capabilities targeting low-latency and on-device use cases.

What It Is For

The stated target verticals: manufacturing and industrial inspection, media content analysis and moderation, robotics and manipulation, geospatial and infrastructure monitoring, and security surveillance. Perceptron describes a second use case gaining traction — augmenting text-first agents like Claude and Codex with a vision layer that handles document parsing, desktop task automation, and browser interactions.

Why the Price Gap Exists

Frontier labs price video and vision models as a premium tier on top of their general-purpose architectures. Perceptron built around a narrower objective: structured output from video, with spatial localization. The tradeoff is 33K context rather than the 1M+ context windows on general frontier models, and no general text reasoning capability. For pipelines that need video intelligence as one component — not as the whole model — that tradeoff is favorable.

The cost-per-task economics for video understanding in production shift materially at $0.30 blended versus $2-3. At 10 million frames analyzed per month, the difference is roughly $27,000 against $200,000-plus.