Thinking Machines Lab Tried to 4x Its Valuation in Months — Here's Why It Couldn't
Most coverage of Mira Murati's AI startup falls into one of two buckets: breathless reporting on the record-setting seed round in 2025, or near-total silence since. Almost nobody connected what happened in between.
Here's the part that's actually interesting: Thinking Machines Lab tried to quadruple its valuation to as much as $60 billion late last year. The deal fell apart. The company is back to where it started, financially — and that's a more useful story than another "AI startup raises billions" headline.
It's also not the only thing that's been quietly happening. There's a new product, a serious talent war, and one investor in the seed round that almost no article has mentioned.
Mira Murati broke 18 months of public silence in June 2026 to reveal Thinking Machines Lab's roadmap — after a valuation episode that tells a more honest story than the headlines did.
What Thinking Machines Lab Actually Is
Thinking Machines Lab is an AI startup founded in February 2025 by Mira Murati, who spent six and a half years as OpenAI's chief technology officer before leaving in September 2024. It's structured as a public benefit corporation, based in San Francisco.
The founding team read like an OpenAI alumni roster: Barret Zoph (former VP of Post-Training Research), Lilian Weng (former VP), and OpenAI co-founder John Schulman, who joined after a brief stint at Anthropic. By launch, the company had hired roughly 30 researchers and engineers pulled from OpenAI, Meta AI, and Mistral.
The company's stated focus is human-AI collaboration — building systems designed around how people actually communicate, rather than optimizing purely for raw model capability.
The Valuation Story Nobody Connected the Dots On
In July 2025, Thinking Machines closed a $2 billion seed round at a $12 billion valuation, led by Andreessen Horowitz with participation from Nvidia, AMD, Cisco, and Jane Street. At the time, it was the largest seed round in AI history.
Four months later, in November 2025, Bloomberg reported the company was in talks for a new round that would value it at $50 to $60 billion — a four-to-five-fold markup in under half a year, on the strength of a single shipped product.
By January 2026, those talks had collapsed. Multiple outlets confirmed the financing never closed; prospective backers weren't willing to support that price without a deeper product track record. The valuation didn't disappear — it simply reset to its original $12 billion seed level.
🔍 The Detail Almost Every Article Has Skipped
Buried in Thinking Machines' seed round is a $10 million investment from the government of Albania — Murati's country of origin. The investment reportedly required an amendment to Albania's 2025 national budget to go through.
There's a second overlooked detail worth knowing if you're tracking who actually controls this company: Murati holds a deciding vote on board matters, and founding shareholders' votes are weighted 100 times heavier than regular shareholders'. That's a notably concentrated governance structure for a company whose CEO has publicly warned about too much consequential AI decision-making sitting in too few hands industry-wide.
Neither detail is hidden, exactly. They're both in public filings and reporting. They're just rarely mentioned in the same article as the funding headlines.
The Actual Product News: Interaction Models
On May 11, 2026, Thinking Machines revealed what it calls "interaction models" — a new AI category built around full-duplex communication. Instead of the turn-based pattern every chatbot uses today, these models process audio, video, and text continuously, in roughly 200-millisecond intervals.
The flagship model, TML-Interaction-Small, responds in 0.40 seconds — close to natural human conversational speed. By comparison, Google's Gemini-3.1-flash-live responds in 0.57 seconds and OpenAI's GPT-realtime-2.0 in 1.18 seconds, according to the company's own published benchmarks.
The company has described its core thesis simply: "Interactivity should scale alongside intelligence." Right now, this is a research preview, not a generally available product — a limited rollout to research partners is planned in the coming months, with a broader public release later in 2026.
⚡ What Makes Interaction Models Technically Different
- Full-duplex processing: generates a response while still receiving input, rather than waiting for a complete turn
- Continuous multimodal input: audio, video, and text processed simultaneously, not sequentially
- 200ms processing intervals: designed to catch interruptions, corrections, and pauses the way a human conversation partner would
- No external control components: the real-time behavior is built into the model itself, not bolted on with separate orchestration logic
The Talent War Behind the Scenes
Meta reportedly attempted to acquire Thinking Machines outright in 2025. When Murati declined, Meta hired away seven of the company's founding members instead, as part of its broader push to staff up its own AI research efforts.
Murati's response was to bring in Soumith Chintala, the creator of PyTorch, as the company's CTO. In January 2026, two more founding-team members, Barret Zoph and Luke Metz, left to return to OpenAI.
At her June 4, 2026 Bloomberg Tech appearance — her first major public appearance in 18 months — Murati addressed the departures directly, framing the volatility as compressed timeline rather than a deeper problem, while declining to dwell on specifics.
🛠️ Tinker — The Product Thinking Machines Already Shipped
- What it is: A fine-tuning API for open-weight language models, released as the company's first commercial product
- Who it's for: Developers and research teams customizing open-source models rather than closed frontier models
- Why it matters: It was the only shipped product on the books when the $50–60B valuation talks began — a key reason those talks ultimately stalled
Where Thinking Machines Actually Stands in Mid-2026
✅ What's Genuinely Working
- Genuine technical novelty with the full-duplex Interaction Models approach
- Heavyweight infrastructure backing, including a gigawatt-scale Nvidia Vera Rubin deal
- Retained its original capital base and core backers despite the stalled markup round
- Still able to attract top-tier talent, evidenced by the Soumith Chintala hire
⚠️ What's Genuinely a Concern
- Lost multiple high-profile founding researchers to OpenAI and Meta within 18 months
- Failed to close a markup round at $50–60B, landing back at its original $12B valuation
- Interaction Models remain a research preview, not a generally available product
- Governance is concentrated, with Murati holding a deciding board vote and founders holding 100x voting weight
What This Actually Means If You're Tracking This Space
💡 Tip #1: Read Valuation Headlines Skeptically From Now On
A company going from $12B to talks at $50–60B and back to $12B, all in six months, on the back of a single product, is a clean case study in how AI funding headlines can outrun product reality. Apply that same skepticism the next time you see a "talks to value at $X billion" headline anywhere in this sector.
💡 Tip #2: If You're Evaluating Tinker, Know Its Actual Scope
Tinker is built for fine-tuning open-weight models, not closed frontier models. If your use case needs the latter, this isn't the tool — set that expectation before you spend evaluation time on it.
💡 Tip #3: Watch the Research-Preview-to-GA Timeline Closely
If you're building real-time voice or video AI products, getting into the limited research-partner rollout for Interaction Models could matter more than waiting for general availability later in 2026 — early access to a genuinely new model category is a real edge.
💡 Tip #4: Track Talent Flows as a Leading Indicator
Founding researchers moving back to OpenAI, and Meta's aggressive poaching, tell you more about where frontier AI talent sentiment is heading than any funding press release. Watch who's hiring whom, not just who's raising what.
✅ Thinking Machines Lab in June 2026 — The Real Picture
- ✅ Founded February 2025 by ex-OpenAI CTO Mira Murati as a public benefit corporation
- ✅ $2B seed round at $12B valuation (July 2025) — the largest seed round in AI history at the time
- ⚠️ $50–60B markup talks collapsed by January 2026; valuation reset to the original $12B
- ✅ Interaction Models revealed May 2026 — full-duplex AI responding in 0.40 seconds, currently a research preview
- ✅ NVIDIA gigawatt-scale Vera Rubin deal confirmed March 2026
- ⚠️ Several founding researchers departed for OpenAI and Meta within the company's first 18 months
- ✅ Albania's government invested $10M in the original seed round
🧮 Startup valuations reset. Where does your personal net worth stand?
Cut through the venture capital noise. Use the free AI Wealth Percentile Engine to instantly benchmark your income and assets against real U.S. economic data.
100% free, no sign-up required.
The Honest Takeaway
Thinking Machines Lab's story in 2026 isn't a straight line up. It's a company that overreached on valuation, got told no, kept its core team and capital largely intact anyway, and shipped something genuinely novel a few months later.
That's a more useful story than either the breathless 2025 coverage or the silence that followed it. Whether Interaction Models becomes a real product category or gets absorbed into existing AI assistants is the open question the rest of 2026 will answer.
For now, the company that tried to be worth $60 billion is worth $12 billion and building something that didn't exist before. That's worth tracking on its own terms.
Frequently Asked Questions
What is Thinking Machines Lab and who founded it?
Thinking Machines Lab is an AI startup founded in February 2025 by Mira Murati, the former chief technology officer of OpenAI, where she worked for six and a half years before leaving in September 2024. The company is structured as a public benefit corporation based in San Francisco, and its founding team included former OpenAI executives Barret Zoph and Lilian Weng, along with OpenAI co-founder John Schulman. It closed a $2 billion seed round in July 2025 at a $12 billion valuation, the largest seed round in AI history at the time.
What products has Thinking Machines Lab actually launched?
As of mid-2026, Thinking Machines has shipped two things: Tinker, a fine-tuning API for open-weight language models that was its first commercial product, and Interaction Models, a new category of full-duplex AI revealed in May 2026 that is currently in research preview rather than general availability. Its flagship interaction model, TML-Interaction-Small, responds in approximately 0.40 seconds, which the company says is faster than comparable real-time models from Google and OpenAI.
What happened to Thinking Machines Lab's $50 billion valuation talks?
In November 2025, Bloomberg reported the company was in discussions with investors at a valuation between $50 and $60 billion, a four-to-five-fold increase from its $12 billion seed valuation just four months earlier. By January 2026, multiple media outlets confirmed those talks had collapsed without a deal, reportedly because prospective backers were unwilling to support that price given the company had only one shipped product at the time. The valuation reset to its original $12 billion level rather than disappearing entirely.
What are "Interaction Models" and how are they different from typical AI chatbots?
Interaction Models are a new AI category Thinking Machines Lab revealed in May 2026, built around full-duplex communication rather than the turn-based, prompt-and-response pattern used by virtually all AI chatbots today. Instead of waiting for a user to finish speaking or typing, these models process audio, video, and text continuously in roughly 200-millisecond intervals, allowing them to register interruptions, corrections, and pauses in something closer to real time. The approach is currently available only as a limited research preview, not a public product.
Why have several Thinking Machines Lab researchers left for OpenAI and Meta?
Several founding-team members have departed since the company's 2025 launch. Meta reportedly attempted to acquire Thinking Machines outright in 2025, and after Mira Murati declined, hired away seven founding members directly. Separately, in January 2026, founding researchers Barret Zoph and Luke Metz returned to OpenAI. Murati has publicly described the departures as a natural consequence of building a frontier AI lab from scratch under intense competitive pressure for talent, rather than addressing the specific circumstances of each departure.
