Former OpenAI chief technology officer Mira Murati, through her new company Thinking Machines Lab, has announced something every AI company is currently chasing - a model that can listen while speaking. The technical term is „full duplex". The concept - artificial intelligence that works like a phone call, not like a chat.

Today every AI model runs on a linear order. The user speaks, the model listens. The model answers, the user listens. An exchange - but not a conversation. TML-Interaction-Small pulls all of this into one. A model that responds in 0.40 seconds - a speed close to natural human dialogue and significantly faster than what OpenAI and Google currently offer.

It is still a research version. Not yet available to the general public. Murati says limited access will open in the coming months, and broader access during 2026.

What is hidden behind this? The idea that interactivity should not be an add-on - but something built into the model itself. The benchmarks Thinking Machines is publishing are striking. But as with everything in the AI industry, numbers under controlled conditions and real-world experience rarely match. The question is not whether „TML-Interaction-Small" can listen while it speaks - it is whether that means anything for an ordinary user, or whether this is yet another demonstration for investors.