Former OpenAI CTO Challenges Old Company: New Model Responds in 200ms, Outperforming GPT-Realtime

According to monitoring by Dongcha Beating, the Thinking Machines laboratory founded by former OpenAI CTO Mira Murati has released a research preview of its “interactive model.” The new system abandons the traditional approach of stitching together voice and text using external tools, instead natively handling real-time audio and video interactions. The model can continuously receive information with a “micro-turn” of 200ms, allowing for simultaneous listening, viewing, and speaking, while supporting real-time interruptions from users. The first showcased model, TML-Interaction-Small, employs a 276 billion parameter MoE architecture, activating 12 billion parameters at a time. To address the traditional large model’s flaw of “stopping perception when generating responses,” the development team has split the system into front-end and back-end: the front-end model is dedicated to maintaining uninterrupted dialogue, while the back-end model simultaneously handles complex reasoning, web searches, or UI generation, seamlessly relaying the results back to the front-end. This architecture directly surpasses the response speed of its old company’s competitors. Official data shows that its voice rotation delay is only 0.40 seconds, achieving a score of 77.8 in FD-bench V1.5, with both core metrics exceeding those of GPT-realtime-2.0 and Gemini 3.1 Flash Live. However, continuous processing of audio and video can quickly deplete context capacity, and the low-latency effect is highly dependent on network conditions. Thinking Machines plans to open a limited preview in the coming months.