Thinking Machines Lab ships its first model and argues interactivity is what OpenAI gets wrong about voice

Thinking Machines Lab has published a research preview of its first AI model, called the Interaction Model. The model is designed to break out of the traditional question-and-answer pattern of voice AI by processing audio, video, and text in parallel 200-millisecond chunks. The startup claims it outperforms OpenAI's GPT-Realtime-2 and Google's Gemini Live on interaction quality. The core innovation is time-aligned micro-turns: the model continuously processes and generates 200ms tokens in an interleaved fashion, eliminating artificial turn boundaries. It uses a fast interaction model paired with a background reasoning model to handle deeper tasks without breaking conversation flow. On the Audio MultiChallenge benchmark, it scores 43.4% — above fast variants of competitors but below GPT-Realtime-2 in thinking mode. Founded by former OpenAI researchers including Mira Murati, Thinking Machines Lab previously raised $2B at a $12B valuation. This model is the first in-house AI product backing Murati's ambition to compete with OpenAI, Anthropic, and Google DeepMind. If successful, it could reshape voice AI by making interactions feel more natural and interruptible, but the company faces execution risks and a rapidly evolving competitive landscape.

ExperimentConfidence: MediumMid term horizonExecution risk: High

Published 12 May 2026, 14:11 UTC4 min read

Visit original article ↗

Summary

TL;DR

Thinking Machines Lab published a research preview of its Interaction Model for voice AI.
The model processes audio, video, and text in parallel 200ms chunks using time-aligned micro-turns.
It claims higher interaction quality than OpenAI's GPT-Realtime-2 and Google's Gemini Live.
The architecture uses a fast interaction model paired with a slower background reasoning model.
On Audio MultiChallenge, scores 43.4% — above fast competitors but below GPT-Realtime-2 thinking mode.
Company founded by former OpenAI researchers and raised $2B at $12B valuation earlier this year.
The release aims to prove interactivity should be native to AI, not a separate harness.

Audience lens

Board

Consider investing in or acquiring interactive AI capabilities to stay competitive.
Monitor Thinking Machines Lab's progress as a potential acquisition target or competitor.

CTO

Evaluate whether to adopt micro-turn architecture for real-time voice applications.
Plan for infrastructure that supports interleaved input/output streaming.

Startup founders

If you are building conversational AI, innovate on interactivity as a differentiator.
Learn from Thinking Machines' paired fast/slow model architecture.

Investors

Assess Thinking Machines Lab's technology if considering investment; weigh execution risks.
Look for other startups applying native interactivity to voice AI.

Policy

Prepare for new regulatory questions around real-time voice AI interactions and data privacy.
Consider implications for synthetic voice authentication.

Engineering

Design systems that can process and generate at sub-second granularity.
Explore time-aligned token streams for multimodal applications.

Developers

Experiment with the Interaction Model if API access becomes available.
Understand how micro-turns differ from traditional request-response patterns.

Content creators

Voice AI improvements could enable more dynamic interactive stories and live translation.
Consider use cases for real-time dubbing or podcast interactivity.

Entrepreneurs

Opportunity to build applications that leverage proactive, interruptible voice AI.
Look for gaps in the market that require natural conversational patterns.

Sales

If your product uses voice, emphasize more natural interactions as a selling point.
Be prepared to explain how you handle user interruptions and proactive prompting.

Marketing

Highlight interactivity as a key feature if adopting voice AI.
Use 'more natural conversations' as a messaging angle.

Accounting

No immediate accounting implications; this is a technology development.

Consumers

Future voice assistants could feel more human-like, with less rigid turn-taking.
Expect more responsive and less 'robot-like' interactions in smart speakers and apps.

Thinking Machines Lab ships its first model and argues interactivity is what OpenAI gets wrong about voice

Summary

Strategic Signals

Audience lens

Sentiment

Impact & reach

Risk radar

Thinking Machines Lab ships its first model and argues interactivity is what OpenAI gets wrong about voice

Summary

Strategic Signals

Audience lens

Sentiment

Impact & reach

Risk radar

Related coverage