Our Take
Murati is signaling product momentum without risking specifics; the 200ms claim is unverified and the release timeline remains vague.
Why it matters
Thinking Machines has operated quietly for 18 months while OpenAI and Anthropic dominated headlines. A public reappearance with a technical demo (even a preliminary one) matters to anyone tracking who is shipping what in foundation models and interfaces.
Do this week
Track Thinking Machines' Tinker API adoption metrics and monitor when 'interaction models' move from preview to production—the gap between announcement and availability will tell you whether the tech is real.
Murati surfaces after 18 months with a new interface model
Mira Murati, CTO of OpenAI for six years and now CEO of Thinking Machines Lab, gave her first major media interview in roughly 18 months on Thursday with Bloomberg. The timing matters: Thinking Machines has spent the past year and a half raising capital, hiring researchers, and shipping Tinker, an API for fine-tuning open-source AI models. During that period, OpenAI and Anthropic have captured most public and investor attention.
Murati used the appearance to preview what Thinking Machines calls "interaction models." Rather than the standard turn-based, prompt-and-response dynamic of most AI products today, these models process continuous streams of audio, text, and video in 200-millisecond intervals. The company frames this as closer to natural human conversation, where interruptions, mid-thought corrections, and pauses are part of the signal. Murati described it as a "fundamentally different kind of AI interface" but was explicit about positioning it as a first step, not a finished product. She declined to announce a specific release date.
The interview also covered Murati's role during OpenAI's November 2023 board crisis, when she became interim CEO after the board fired Sam Altman. She said she felt clear about her decisions to protect the mission and team, and claimed the company would have "imploded" without her involvement. Asked directly whether she still trusts Altman, she sidestepped, pivoting instead to concerns about concentration of power in AI governance. She suggested the industry has focused too much on the virtue of individual leaders and too little on structural checks.
When pressed on departures of high-profile researchers from Thinking Machines, Murati downplayed the issue, attributing volatility to building a frontier lab from scratch and acknowledging that compensation (nine-figure packages in the AI talent war) matters but isn't the whole story.
Visibility is currency in a crowded AI market
Thinking Machines operates in a landscape where attention and capital flow to the noisiest competitors. OpenAI is constantly in the news cycle. Anthropic's momentum dominates current conversation. xAI was folded into SpaceX with expected significant public offering coverage. In that environment, the source notes, "staying heads down has diminishing returns; at some point, you have to make some noise just to remind the market you exist."
Murati's interview serves that function without overcommitting. She has previewed a technical direction—200ms interaction loops—without publishing benchmarks, independent verification, or a concrete timeline. The move is calibrated: enough detail to signal technical work, enough caution to avoid claims that could later be walked back. For practitioners evaluating foundation model vendors and interfaces, the absence of independent reproduction or customer deployment examples means the interaction model remains a prototype claim, not a shipping product.
The governance angle Murati raised—structural checks on AI decision-making—is a separate signal about how she is positioning Thinking Machines as distinct from the personality-driven narratives of OpenAI and Anthropic. Whether that framing attracts researchers, customers, or investors remains unseen.
Separate the signal from the specification
Murati's 200-millisecond processing claim is interesting but unverified. No independent benchmarking, no customer deployment, no published paper. The interaction model concept (streaming multimodal input) is not novel in research; the competitive claim is the latency and continuity. Before allocating time or budget to evaluate Thinking Machines as a vendor alternative, wait for either independent reproduction of the 200ms claim or a customer case study showing real-world advantage over turn-based systems. Tinker (the fine-tuning API) is already shipping and may offer differentiation on cost or usability; focus evaluation there if you are considering open-source model tuning. Do not confuse Murati's careful media appearance with product readiness.