What are AI interaction models?

Interaction models are a two-tier architecture for real-time human-AI collaboration: a foreground model continuously processes input and output in 200-millisecond micro-turns, handling audio, video and text together, without waiting for the user to finish speaking before starting to respond; a background model works asynchronously on deeper reasoning, tool use and long-horizon tasks, weaving results into the ongoing conversation as soon as they are ready, without blocking the real-time interaction. The term was coined by Thinking Machines Lab, the lab founded by Mira Murati, formerly OpenAI's CTO, which in May 2026 introduced the first model in this category, TML-Interaction-Small, a large mixture-of-experts model built specifically for this kind of real-time exchange. The difference from traditional conversational models, which wait for a complete input before generating a complete response, is structural: here perception and generation happen in the same continuous stream, not in two separate phases.

See also: AI Agents (Agentic AI) · AI systems that go beyond answering: they plan, use tools and take actions autonomously inside your processes.

The bottleneck it solves

Rigid turn-based conversational systems impose a limit that has nothing to do with how intelligent the model is and everything to do with the shape of the interaction itself: the user must finish the thought, the system must finish the answer, and nothing can happen in between. There is no natural interruption, no reacting to hesitation in someone's tone, no correcting mid-sentence the way a person would in a real conversation. Thinking Machines Lab calls this limit the "collaboration bottleneck": the more capable the model becomes, the more this turn-based rigidity becomes the actual constraint, because the bandwidth of communication between human and AI stays that of a back-and-forth interface regardless of how sophisticated the model behind it is.

A concrete enterprise example

A clear application is a voice agent for customer service: with a turn-based model, the agent waits for the customer to finish speaking, then processes and emits the full response, even if the customer is already showing frustration halfway through the sentence. With an interaction model, the foreground layer picks up the tonal shift in real time and can adapt its register or offer an escalation before the customer finishes expressing the problem, while the background model checks the account history or processes a refund in parallel, without making the conversation wait. This is not computer use, which is about an agent operating a graphical interface on a person's behalf: here the subject is the conversational channel itself, how audio, video and text are exchanged in real time between human and AI.

Why it matters, with the caution it deserves

For anyone evaluating a customer experience or agentic CX use case with voice or video interaction, this paradigm promises to remove the perceived latency and turn-based rigidity that today give away the artificial interlocutor. But it needs to be treated for what it is right now: TML-Interaction-Small is available only in a limited research preview for a select set of partners, with no general public release announced yet, and other labs such as ElevenLabs appear to be independently converging on the same category language, a sign the paradigm is emerging rather than an established market standard. Building a production use case on top of it today means depending on restricted access and an API that has not yet stabilized: the sensible call for now is to monitor how it evolves and consider a pilot once access widens, not to plan a production launch on the announcement alone.