iRTE 2026: A Look at the Real-Time Models OpenAI and Thinking Machines Are Betting On

The Chinese publication Jiqizhixin (机器之心) has announced a discussion of real-time interaction models at the iRTE 2026 conference. According to the organizers, OpenAI and Thinking Machines Lab are betting on this direction. Over the past year, multimodal teams have stopped optimizing for a "more human-sounding voice response" and taken on a harder problem: how AI takes part in an interaction that is unfolding right now. The available excerpt of the announcement includes no program, speakers or dates, but the framing of the question itself shows where the focus has shifted.

Three stages of one problem

The organizers describe the progression as a ladder with three rungs:

• Full-duplex models. The agent listens and speaks at the same time. It can be interrupted, and it does not mistake a pause in speech for the end of a turn.
• Full multimodality (omni). Voice, vision and other signals are processed together, so decisions rest on the whole picture rather than on a transcript of speech.
• Interactive models. The next step: understand what exactly is happening in the conversation at this second and choose the next action.

Each stage adds a new layer of responsibility on top of the previous one. Duplex handles timing: when to come in and when to stay quiet. Omni handles perception: what the model sees and hears. An interactive model handles the meaning of the situation: who is leading right now, what the other person expects, whether it is appropriate to step in.

Why the bet on real time looks logical

First-generation voice assistants worked on a "hear, transcribe, think, speak" pipeline. Any such chain adds latency and breaks down on live speech, where people interrupt each other, correct themselves and say "well..." instead of answering. The synthesized voice can sound great, yet the conversation still feels like swapping notes.

OpenAI has already brought voice modes and the Realtime API into its products. Thinking Machines Lab, the startup of former OpenAI CTO Mira Murati, is looking in the same direction, according to the organizers. If two teams of this caliber independently arrive at the same vector, that is a signal to the market. Real-time conversation is becoming a distinct class of problem with its own metrics and architecture, not a layer on top of a text LLM.

What the announcement leaves out

The announcement contains no benchmarks, no model names and no timelines. The claim that "both companies are betting on this" comes from the conference organizers, and the text includes no confirmation from the companies themselves. The term "interactive model" does not yet have a single definition: the organizers do not say where the line falls between a strong omni agent and a model that understands the course of an interaction. Judging by the stated topic, the substantive discussion of exactly this will take place at the conference itself.

What teams building voice agents should check

If you are building a voice or video agent, this shift has a practical consequence: you will have to revisit your quality metrics. Recognition accuracy and voice naturalness alone are no longer enough. It is worth measuring:

• the delay between the end of the user's turn and the start of the response;
• how correctly the agent reacts to interruptions: whether it stopped talking and kept the context;
• the rate of false starts, where the agent spoke while the person was simply thinking;
• the ability to use visual context when it is available: screen, camera, gestures.

It is also worth deciding up front what you are building: a cascade of separate ASR, LLM and TTS components, or a single model that works with audio directly. Both approaches coexist today, but the conversation at iRTE 2026 shows where the top end of the market is heading.