Perceptron Inc. on Tuesday launched its flagship video analysis AI model Mk1, priced at $0.15 per million input tokens and $1.50 per million output tokens — roughly 80-90% cheaper than competing models from Anthropic, OpenAI, and Google. The two-year-old startup’s model specializes in understanding live video feeds and analyzing visual content for enterprise applications.
According to Perceptron’s announcement, the model was developed over 16 months by a team led by CEO Armen Aghajanyan, formerly of Meta FAIR and Microsoft. The company designed what it calls a “multi-modal recipe” from scratch to handle real-world video analysis tasks.
Video Analysis Capabilities and Use Cases
Mk1 targets enterprise applications requiring real-time video understanding. The model can function as a security monitoring system for facilities, automatically clip highlights from marketing videos for social media repurposing, and identify inconsistencies in video content for quality control.
Additional use cases include analyzing body language and participant actions in controlled studies or job interviews. VentureBeat reported that while some AI models offer video analysis today, it remains far from mainstream capability.
The model demonstrates understanding of cause-and-effect relationships, object dynamics, and physical laws — capabilities the company positions as essential for real-world video analysis applications.
Thinking Machines Previews Real-Time Interaction Models
Meanwhile, Thinking Machines announced a research preview of “interaction models” designed to move beyond turn-based AI conversations. The startup, founded by former OpenAI CTO Mira Murati and researcher John Schulman, treats interactivity as a core architectural component rather than external software.
These models can process and respond to inputs while simultaneously handling new incoming data across text, audio, and video formats. The technology aims to enable more natural, fluid human-AI interactions for applications requiring real-time collaboration.
Thinking Machines plans to open a limited research preview in the coming months before broader availability. The models are not yet accessible to public or enterprise users.
Sakana’s RL Conductor Orchestrates Multiple AI Models
Sakana AI introduced RL Conductor, a 7-billion parameter model trained via reinforcement learning to automatically coordinate multiple large language models including GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro. The research paper shows the system dynamically analyzes inputs and distributes tasks among different AI workers.
Yujin Tang, co-author of the research, told VentureBeat that manual frameworks like LangChain “start breaking the moment the query distribution shifts.” RL Conductor addresses this by automatically adapting coordination strategies based on input patterns.
The system achieves state-of-the-art performance on reasoning and coding benchmarks while using fewer API calls than traditional multi-agent pipelines. Sakana commercialized the technology through Fugu, its multi-agent orchestration service.
Performance and Cost Benefits
RL Conductor outperforms individual frontier models and expensive human-designed systems while operating at a fraction of the cost. The automated approach eliminates the need for manual pipeline coding that typically breaks when user demands shift.
Tang noted that production systems serving large, diverse user bases require “going beyond human-hardcoded designs” to achieve real-world generalization across heterogeneous applications.
Musk-OpenAI Trial Continues with Key Testimonies
The high-stakes trial between Elon Musk and OpenAI entered its third week with testimony from Microsoft CEO Satya Nadella and OpenAI co-founder Ilya Sutskever. The Verge reported that Sam Altman testified Tuesday to refute Musk’s characterizations.
Musk’s 2024 lawsuit accuses OpenAI of abandoning its founding mission to develop AI for humanity’s benefit in favor of profit maximization. Previous witnesses included Musk himself, Neuralink CEO Jared Birchall, and former OpenAI board member Shivon Zilis.
The trial could significantly impact OpenAI’s future structure and ChatGPT’s development trajectory. OpenAI maintains that Musk’s claims misrepresent the company’s evolution and original agreements.
What This Means
These developments signal three major shifts in AI model deployment. First, specialized models like Perceptron’s Mk1 are competing on cost efficiency rather than general capability, potentially democratizing access to advanced video analysis. Second, real-time interaction models from Thinking Machines could eliminate the turn-based limitations that currently constrain AI applications.
Most significantly, Sakana’s RL Conductor demonstrates that smaller models can effectively orchestrate larger ones, potentially reducing reliance on single frontier models. This approach could lower costs while improving performance through automated coordination.
The ongoing Musk-OpenAI trial adds regulatory uncertainty to the landscape, with potential implications for how AI companies structure their missions and profit models.
FAQ
How much cheaper is Perceptron Mk1 compared to other video AI models?
Perceptron Mk1 costs $0.15 per million input tokens and $1.50 per million output tokens, which is 80-90% less expensive than comparable models from Anthropic, OpenAI, and Google.
What makes Thinking Machines’ interaction models different from current AI?
Unlike current turn-based AI systems, interaction models can process and respond to inputs while simultaneously handling new incoming data, enabling more natural real-time conversations across text, audio, and video.
How does Sakana’s RL Conductor improve on existing multi-agent AI systems?
RL Conductor automatically coordinates multiple AI models through reinforcement learning, eliminating the need for manual pipeline coding that breaks when user demands change. It achieves better performance with fewer API calls than traditional approaches.
