Thinking Machines Lab is finally showing more of its hand, and the pitch is clear: make AI feel less like prompting software and more like working with another person in real time.
Mira Murati's post-OpenAI company is no longer just a symbol of the talent spillover from the industry's most famous AI lab. Thinking Machines Lab is now demonstrating what it calls interaction models, a new class of AI systems built to continuously take in audio, video and text while responding in real time.
That matters because the next fight in AI is not only about who has the largest model or the most polished chatbot. It is about who can define the interface. OpenAI has ChatGPT, Anthropic has Claude, Google DeepMind has Gemini, and Meta has open models and distribution. Murati's company is trying to claim a different opening: AI that can stay present through a live session, notice what is happening, interrupt when useful and keep working while the user is still talking or showing something.
According to The Verge's Jay Peters, Thinking Machines is preparing a limited research preview in the coming months and is aiming for a wider release later this year. The examples are deliberately simple, but revealing. The model can listen for animal mentions in a story, translate speech as it happens, and warn a user when they are slouching. None of those demos, by itself, sounds like a company worth billions. Together, they point to a product thesis that today's AI interfaces are too narrow for the way people actually work.
Thinking Machines says current models are mostly turn-based. A person finishes typing or speaking, then the model responds. If the model is generating, its view of the world effectively pauses until that output ends or is interrupted. That may be acceptable for drafting an email, but it is awkward for tutoring, coding, design review, live translation, robotics-adjacent work, or any task where timing and context matter.
The company's proposed answer is a model that treats interaction as native, not as a layer bolted on through speech detection, transcription, text generation and voice output. Its research post describes a time-aware system that works in small slices, continuously processing streams rather than waiting for neat conversational turns. It also separates fast interaction from deeper background reasoning, so a user can keep talking while another part of the system handles tool use, search or longer-horizon work.
This is the strongest evidence that Thinking Machines is not only trading on Murati's name. The company has already released Tinker, an API for fine-tuning open models, and now it is sketching a broader direction around customization and live collaboration. The question is whether those pieces become a coherent product line or remain impressive research artifacts in a market where users already have access to real-time voice modes from OpenAI and Google.
Murati has credibility because she helped lead OpenAI's work on ChatGPT, DALL-E and other systems that shaped the modern AI market. Thinking Machines also launched with serious technical weight, including OpenAI co-founder John Schulman as chief scientist and Barret Zoph as its original CTO. But founder pedigree is not a moat by itself. Zoph later left, Soumith Chintala became CTO, and the company has faced departures to OpenAI and Meta. In frontier AI, talent is both the asset and the battlefield.
Capital is part of the product
The funding signals are as important as the demos. Thinking Machines has reportedly raised about $2 billion, with investors including Andreessen Horowitz and strategic backers tied to the chip ecosystem. Nvidia's multiyear partnership, which includes access to Vera Rubin systems starting in 2027 and at least one gigawatt of compute, puts the company in a category most startups cannot enter.
That is the uncomfortable truth about the post-OpenAI founder ecosystem. The best-known alumni can form venture-scale companies before the product is fully visible because the market is not only buying revenue. It is buying a credible path to talent, compute, enterprise trust and future model capability. A normal startup proves demand, then raises to scale. Frontier AI often works in reverse: raise enormous capital, secure infrastructure, recruit a lab, then make the product legible.
For rivals, this creates a strange competitive map. OpenAI and Anthropic have mature consumer and enterprise brands. Google DeepMind has research depth and distribution through Google. Meta can spend heavily while pushing open models into the developer market. Thinking Machines has to compete without the same installed base, which means it needs a sharper wedge. Interaction models could be that wedge if they make AI more useful in work that is continuous, visual, spoken and messy.
The risk is that everyone else can see the same problem. OpenAI, Google and Anthropic are all working toward richer multimodal agents, and Meta has the resources to turn research concepts into broadly available developer tools. If Thinking Machines proves the interface works, larger labs may move quickly. The startup's advantage will have to come from execution speed, model quality, developer adoption and a clear answer to why users should build around its platform instead of waiting for the giants.
Murati's company is now easier to understand, but not easier to judge. It is trying to make the AI assistant less transactional and more collaborative, while building the capital base needed to stay in the frontier race. The next thing to watch is not only when the preview arrives, but whether developers and early customers feel a real difference once the demos leave the stage.
Also read: The EU is turning AI cyber models into a trust test • An Optane home server makes trillion parameter AI feel almost practical • Meta's AI copyright fight now reaches Mark Zuckerberg personally
