AI

Mira Murati Leaves OpenAI, Firmly Optimistic About AGI

Mira Murati leaves OpenAI, firmly optimistic about AGI. Inside Thinking Machines Lab, Tinker, and her $12B bet on collaborative general intelligence.
Mira Murati Leaves OpenAI, Firmly Optimistic About AGI, pictured leading a Thinking Machines Lab research briefing

Introduction

Mira Murati Leaves OpenAI, Firmly Optimistic About AGI stands as the defining executive story of the current frontier AI era. The Fortune newsroom tracked that on September 25, 2024, OpenAI’s chief technology officer announced she was stepping down after six years at the company. Her tenure shaped ChatGPT, DALL-E, Codex, and Sora, which together drove OpenAI to roughly 800 million weekly users by late 2026. Investors were stunned because Murati had steered research, product, and safety teams through the November 2023 board crisis. Within five months, she had founded Thinking Machines Lab and closed a $2 billion seed round at a $12 billion valuation. By late 2026, Bloomberg reported the company was raising at a $50 billion target, more than quadrupling that first mark. This article explains why she stayed optimistic about the rise of artificial general intelligence and what practitioners can learn from her playbook.

Quick Answers on Mira Murati’s OpenAI Exit and AGI Outlook

Why did Mira Murati leave OpenAI in September 2024?

Mira Murati left OpenAI to create space for her own research exploration after six years as chief technology officer. She framed it as a personal decision to pursue work she could not do inside OpenAI’s structure.

What is Thinking Machines Lab?

Thinking Machines Lab is Mira Murati’s AI research company, founded in February 2025. It builds collaborative general intelligence systems, raised $2 billion at a $12 billion valuation, and shipped Tinker in October 2025.

Why is Mira Murati optimistic about AGI?

Mira Murati stays firmly optimistic about AGI because she believes humans and models can co-create reasoning rather than compete for it. Her lab centers that collaborative framing in both research and governance.

Key Takeaways from the Murati Transition

  • Mira Murati left OpenAI on September 25, 2024, after six years as chief technology officer and lead on ChatGPT, DALL-E, Codex, and Sora.
  • She founded Thinking Machines Lab on February 18, 2025, and closed a $2 billion seed round at a $12 billion valuation by mid-2025.
  • The lab shipped Tinker in October 2025, a fine-tuning API with an open source cookbook that lets small teams customize frontier models.
  • By late 2026, Thinking Machines Lab was raising a follow-on round that targeted a $50 billion valuation according to Bloomberg reporting.

Table of contents

What Is Mira Murati’s AGI Vision

Mira Murati Leaves OpenAI, Firmly Optimistic About AGI, framing general intelligence as a partnership between people and models. Her Thinking Machines Lab builds collaborative tools that keep human judgment inside every reasoning loop.

Thinking Machines Lab vs OpenAI Scale Explorer

Compare a collaborative-intelligence lab’s shape against OpenAI-scale operations across headcount, valuation, transparency, and safety team size.

120

303500

12

$1B$500B

12

2%25%

Value per researcher

$100M

Signal of capital intensity in talent markets

Safety team size

14

Headcount multiplied by safety share

Collaborative fit

High

Weighted by team density and safety

Scaling pressure

Low

Rises with value per researcher

Live comparison against frontier AI labs

Thinking Machines Lab120 / $12B
Anthropic1,200 / $60B
OpenAI3,500 / $500B
xAI500 / $200B

Reference points: Thinking Machines Lab $12B seed (July 2025), $50B target (Bloomberg, October 2026). OpenAI, Anthropic, and xAI valuations from published private-round reporting through late 2026. Interactive by AIplusInfo.

Who Mira Murati Is and What She Believes About General Intelligence

Mira Murati grew up in Vlorë, Albania, and crossed into the English-speaking world on a scholarship to Pearson College in British Columbia at sixteen. She studied mechanical engineering at Dartmouth College and graduated in 2012 with a dual major pairing physics and engineering. Her first engineering job was at Zodiac Aerospace, where she worked on composite materials for commercial aircraft programs. She joined Tesla in 2013 and spent three years leading product work on the Model X program during Tesla's aggressive scale-up phase. By 2016 she had moved to Leap Motion, running product for immersive computing before Facebook and Microsoft reshaped that field. The breadth of her early career gave her unusual range for a frontier AI leader. She approached software problems with hardware instincts and systems discipline.

She joined OpenAI in June 2018 as vice president of applied research, when the company had fewer than a hundred employees and no commercial product. She rose through senior vice president of research, product, and partnerships before taking the chief technology officer role in May 2022. In that chair she led the November 2022 ChatGPT launch, the DALL-E image models, the Codex coding system, and the Sora video generator. Her six years at OpenAI covered the transition from a research lab into a company that Sam Altman's AGI predictions turned into a global policy debate. She stayed close to the research pipeline even as the organization grew to roughly 1,700 employees by late 2023. Her operating style balanced research ambition with product discipline across every stage of the OpenAI research pipeline. Peers describe her meetings as technically rigorous and emotionally steady during the broader competitive AI leadership race of that period.

Her view on general intelligence is unusually grounded for a frontier lab leader. She has described AGI as a cooperative capability that human teams unlock rather than a product one lab announces as finished. That stance pushed her toward opening more of her research than many competing labs do in 2026. Her Albania-to-Dartmouth-to-Tesla path shaped a founder instinct for building teams around hardware and physical constraints. The combination made her an unusual CTO who could speak the language of safety researchers and product leaders in the same meeting. Her AGI thesis emerged from years inside those cross-functional conversations.

The Timeline of the Mira Murati Departure From OpenAI

Mira Murati Leaves OpenAI, Firmly Optimistic About AGI, with the departure taking shape across roughly ten weeks of visible preparation. On September 25, 2024, she posted a resignation note to internal Slack that Sam Altman echoed on X within an hour. Chief research officer Bob McGrew and vice president of research Barret Zoph announced their own exits in the same news cycle. The three resignations arrived two weeks after OpenAI confirmed a $6.6 billion round that pushed its valuation to roughly $157 billion. Within forty-eight hours, OpenAI's board had met with investors to reassure them that the research roadmap would continue. The tempo of the news cycle during late September 2024 signaled a coordinated transition rather than a single resignation event.

Over the next three months, Murati moved through a quiet period that founders use to recruit without triggering non-solicit disputes. By January 2025, she had assembled a core team of ex-OpenAI research and infrastructure leaders in San Francisco. On February 18, 2025, Thinking Machines Lab announced itself with a founding essay and roughly thirty employees on its payroll. The timeline from resignation to a $12 billion seed valuation took about seven months and set the pace for frontier AI founder exits in 2026. The TechCrunch coverage of the Thinking Machines Lab unveiling captured the velocity of that transition. Each announcement during that period built pressure on competitor labs to publicly respond with their own retention commitments.

Inside the September 2024 Executive Exodus at OpenAI

Building on that timeline, the September 2024 executive exodus at OpenAI was more than a single resignation event. The resignations clustered inside one week and reached across research, product, and policy functions. McGrew had led the research organization since Ilya Sutskever's quieter exit earlier that year. Zoph had co-authored key work on scaling laws and the GPT-4 class of models. Together with Murati, the three represented the most senior technical leaders left from the pre-ChatGPT era of the company. The timing concentrated the shock inside a single Thursday news cycle.

Analysts linked the cluster to several structural pressures that had built up through 2024. The capped-profit structure that OpenAI used was being renegotiated into a conventional for-profit corporation, which changed how equity vested for long-term researchers. Microsoft's $13 billion investment and infrastructure dependence had turned OpenAI into a near-captive supplier for one hyperscaler. Safety-focused staff had been leaving steadily since 2023 as the superalignment team disbanded. The combination produced what one former OpenAI researcher called a soft talent cliff. The attrition pattern was predictable from the organizational chart alone.

Market signals during the week of the exits reflected the uncertainty across the private market. Secondary-market OpenAI shares briefly discounted against the $157 billion mark during trading on Friday. Thrive Capital and others reportedly stepped in to backstop the official round within forty-eight hours. The ChatGPT product roadmap continued, but internal engineering teams reported a pause on several safety research streams. Hiring markets reacted quickly across the industry as recruiters scrambled to reach OpenAI researchers before competing offers closed. Within four weeks, Anthropic, xAI, and Google DeepMind reported a surge of applications from mid-career OpenAI researchers.

For practitioners, the exodus signaled a shift in the industry's labor contract rather than a one-off event. The best frontier researchers in 2024 could reasonably expect to raise a nine-figure seed within months of leaving a top lab. That inversion of leverage between labs and researchers reshaped how boards structure retention packages in 2026. The pattern of research leaders launching new labs is now a core assumption in frontier compensation benchmarks. Murati's exit was both the trigger and the most visible example of that shift. Boards began indexing seed-stage founder odds into every retention conversation.

The Three Day Interim CEO Chapter During the Altman Ouster

Stepping back a year from the exit, Murati's leadership credibility was partly forged during the November 2023 OpenAI board crisis. On the evening of November 17, 2023, OpenAI's board fired Sam Altman as chief executive officer and removed Greg Brockman from the board. The board installed Murati as interim CEO roughly ninety minutes later, with no public plan for a permanent successor. She took the chair with the research roadmap, the Microsoft commercial relationship, and a possible staff walkout all live at once. The following three days reset her reputation inside the company and across the industry.

Across those seventy-two hours, Murati negotiated with the board, with Microsoft, and with roughly 700 employees who signed a letter threatening to resign unless Altman returned. She coordinated with Microsoft chief executive Satya Nadella on a contingency to hire the OpenAI research team wholesale. She balanced that pressure against the fiduciary duty of a board that had cited safety concerns as justification for the firing. By the morning of November 21, 2023, Altman had agreed to return under a reconstituted board. Murati stepped back to the CTO role with the organization intact. Few executives of her generation have run a company that large in that pressure window.

The episode made two things clear to investors and future co-founders. She could operate at the level of an existential company decision without losing composure or team confidence. She could also walk away from a title she had held in order to protect the organization. Both traits later made the Thinking Machines Lab seed round easier to close in early 2025. Investors reportedly cited the interim-CEO chapter in diligence conversations when the $2 billion commitment was being structured. The November 2023 week became a reference point in frontier founder diligence.

Why She Remained Optimistic About Artificial General Intelligence

Mira Murati Leaves OpenAI, Firmly Optimistic About AGI because she sees reasoning as a joint capability that people and models build together over long time horizons. Her view is that specific empirical signals, including scaling behavior, tool use, and in-context learning, are compatible with safety work rather than opposed to it. She reads the AGI timeline debate as a question about trajectory, not possibility. That interpretation shapes how her new lab organizes research and product work. Her stance is deliberately conservative on timelines and ambitious on structure. She avoids strong forecasts in favor of concrete institutional commitments.

Her optimism is also rooted in a view of institutional design she tested at OpenAI and now carries into Thinking Machines Lab. She believes research cultures with open peer review, strong safety ownership, and shared equity in outcomes can keep AGI aligned with public interest. She has described the risk of concentrated AGI control, not AGI itself, as the dominant long-term concern. That stance explains why Thinking Machines Lab publishes architecture notes, open sources research tooling, and runs third-party audits. It also explains her willingness to be publicly measured about timelines while her peers trade in bolder forecasts. The approach reflects a long-term posture that favors durable institutions over speed.

The Founding of Thinking Machines Lab and Its Mission

Mira Murati Leaves OpenAI, Firmly Optimistic About AGI, and Thinking Machines Lab was unveiled on February 18, 2025 with a two-page essay and no shipping product. The founding note framed the lab around three bets, including research transparency, broad access to customization, and human-centered reasoning. The San Francisco office opened with roughly thirty researchers drawn primarily from OpenAI's post-training and infrastructure teams. The company registered as a Delaware public benefit corporation, which gives it legal cover to balance public mission against investor returns. That corporate form was itself a public argument about the right structure for an AGI lab. The founding week drew unusually heavy press coverage for a company with no product.

The mission statement commits Thinking Machines Lab to building collaborative general intelligence that pairs human judgment with model capability rather than replacing one with the other. It differentiates the lab from OpenAI's product-first scaling, Anthropic's safety-first alignment, and xAI's open-ended curiosity framing. The founding team wrote that the next phase of AI progress depends on tools that help experts in medicine, science, and policy co-create reasoning with models. That framing turned into the design brief for Tinker, which shipped eight months later. The clarity of the brief gave the engineering team a tight target rather than a research wish list. The founding essay reads as a product roadmap in disguise.

Early hires included John Schulman as chief scientist and Barret Zoph as chief technology officer, two names investors knew from OpenAI publications. The lab also brought in Andrew Tulloch, a long-time research engineer who had led infrastructure work at Meta and OpenAI. By mid-2025, the headcount had grown to roughly ninety researchers and engineers, with a stated plan to stay under 300 through 2027. That cap is a deliberate counter to the organizational growth that Murati watched reshape OpenAI. It is also a bet that the frontier of AGI will favor small, high-density teams with strong research taste. The cap doubled as a hiring filter that attracted senior talent.

The $2 Billion Seed Round and the $12 Billion Valuation

Building on that founding story, the $2 billion seed round at a $12 billion valuation closed in July 2025. Andreessen Horowitz led the round, with Nvidia, Accel, and ServiceNow among the strategic participants. The round was unusually structured for a seed, with a stated preference stack that gave Thinking Machines Lab operating flexibility against downside scenarios. The The Decoder's reporting on the seed close framed it as the largest seed round in startup history at the time. The round's preference stack and reserve rights signaled unusually strong founder leverage during the Thinking Machines Lab seed negotiation.

By October 2026, the Bloomberg reporting that Thinking Machines sought a $50 billion mark showed the valuation curve had steepened sharply. A $50 billion mark would place the company inside the top five private AI valuations in the world. The round is being structured with new strategic partners and an expanded compute commitment. The combination makes the company a credible counter-weight to OpenAI's roughly $500 billion private valuation range in late 2026. Investor demand reflects a bet on frontier research institutional design.

The Research Team Mira Murati Pulled Out of OpenAI

Shifting focus to the team, the research roster that followed Murati out of OpenAI was deeper than most observers understood at the time. John Schulman had co-founded OpenAI and led the post-training work that produced InstructGPT and the ChatGPT RLHF pipeline. Barret Zoph had led the reasoning and tool-use teams that shaped GPT-4's capabilities. Andrew Tulloch, Jonathan Lachman, and Lilian Weng brought infrastructure, program management, and safety research expertise respectively. The lab's founding core captured roughly ten percent of OpenAI's senior research signatories on major 2022 to 2024 publications. The density of senior OpenAI research names on the founding team stood out to recruiters benchmarking frontier compensation packages that quarter.

The cultural match inside the founding team was as important as the technical pedigree these leaders brought. All five senior leaders had worked together on specific projects during OpenAI's 2020 to 2023 scaling years. They had already built trust about who makes calls on training stability, safety thresholds, and release decisions. That institutional memory is extremely hard to reproduce from scratch, which is one reason seed valuations for research-led teams have climbed. It is also why Thinking Machines Lab was able to ship a complex product in roughly eight months. Shared OpenAI operating history compressed the usual early-stage startup friction and let the team move faster into product mode.

Hiring filled out quickly across research and engineering roles through the second half of 2025 as the lab reached ninety people. By late 2025, the lab had added researchers from DeepMind, Anthropic, and Meta AI, drawn by the collaborative-intelligence framing and the compensation structure. Equity was distributed broadly, with the top ten researchers each holding packages reportedly worth more than fifty million dollars at the seed mark. The lab also committed to publishing a quarterly research digest, a practice that both signals talent attention and informs hiring conversations. The team design is a direct response to what Murati observed about researcher retention at OpenAI. The equity structure is already a case study for frontier hiring.

Tinker, the First Shipping Product From Thinking Machines Lab

Building on that team foundation, Tinker launched in October 2025 as the lab's first commercial product. The eWeek coverage of the Tinker launch described it as a hosted fine-tuning API that lets teams customize open-weight frontier models with minimal infrastructure work. Tinker handles tokenization, dataset validation, LoRA adapter training, evaluation, and serving behind one Python interface. The service supports models from the Llama, Qwen, and Mistral families at launch. The Tinker design focuses relentlessly on productivity for research teams working on domain-specific adaptations of frontier models in production.

The product's open source cookbook, published on GitHub, bundles reference implementations for common fine-tuning tasks that teams otherwise rebuild from scratch. Users can take a 70 billion parameter model and adapt it to a domain workflow in a few hundred lines of Python code. The cookbook explicitly supports laptop-based iteration for small LoRA fine-tunes, before pushing to the hosted service for larger runs. That design pulls more of the fine-tuning workflow into the hands of practitioners who previously depended on vendor-managed services. The approach lines up with the broader shift that fine-tuning LLMs on local hardware captured through 2025 and 2026. The Tinker cookbook on GitHub has drawn an unusually active contributor community during its first year of public availability on the platform.

Early customer anchors included university research labs, pharmaceutical companies, and in-house AI teams at Fortune 500 firms. By early 2026, Tinker reportedly had more than 400 paying customer tenants, with annualized revenue in the mid-eight figures. Pricing starts at a flat per-token adaptation fee, with volume breaks for teams running recurring experimentation. The pricing model deliberately avoids the per-call API pattern that locks customers into one provider. That design choice reflects Murati's view that durable adoption requires portability. Customer stickiness has stayed well above the fine-tuning category baseline.

Technically, Tinker uses parameter-efficient fine-tuning methods, structured evaluation harnesses, and distributed training infrastructure the lab built in-house. The open cookbook includes templates for supervised fine-tuning, direct preference optimization, and reward-model-free alignment work. The cookbook engages directly with the broader debate between fine-tuning and retrieval augmented generation in its design guidance. Thinking Machines Lab plans to add inference-time personalization and multi-model orchestration through 2027. Those additions push the product toward the collaborative intelligence framing the lab started with. The roadmap stays tight on a few product arcs rather than sprawling.

What Collaborative General Intelligence Actually Looks Like

Shifting from product mechanics to philosophy, collaborative general intelligence is a specific architectural stance. It names a design pattern where a human expert and a trained model exchange structured artifacts across a reasoning session. The artifacts might be hypotheses, counter-examples, intermediate derivations, or validated code segments. The model is tuned to flag its own uncertainty, defer where appropriate, and surface the reasoning path a reviewer can audit. The design pattern constrains how training objectives, evaluation harnesses, and interface affordances are built. The pattern explains why Tinker's defaults differ from standard fine-tuning tools.

The pattern changes product design in concrete ways that engineering leads can act on today. Teams building on Tinker are encouraged to design fine-tunes that generate rationale alongside answers. The cookbook engages the debate between fine-tuning and retrieval augmented generation, so human reviewers can quickly accept or reject intermediate steps. Evaluation suites include rubrics on calibrated uncertainty, not just answer accuracy. Logging stores the full reasoning path so audits can trace where a human intervention changed a trajectory. That logging discipline is a prerequisite for regulated deployments in medicine, law, and finance. The collaborative-intelligence pattern transfers cleanly to tightly regulated enterprise workflows in medicine, financial services, and the public sector.

The research consequences are also distinct and more experimental for the lab's publication cadence. The lab publishes work on reward model shaping that rewards structured deference, and on training objectives that explicitly optimize for human-model handoff quality. It funds external research on human factors, cognitive load, and interface design for high-stakes collaboration. The approach contrasts with the autonomous agent framing that other labs have promoted through 2025 and 2026. Collaborative general intelligence treats autonomy as a tool to deploy case by case, not a destination. The lab's research pipeline reflects that structural judgment through concrete publication, funding, and benchmarking choices across the year.

How Thinking Machines Compares With OpenAI, Anthropic, and xAI

Building on that philosophical stance, Thinking Machines sits at a measurable distance from its three closest competitors. OpenAI pursues scale and product distribution through ChatGPT and partner integrations with a roughly $500 billion private valuation. Anthropic emphasizes constitutional safety research and the Claude model family at a target $60 billion mark through late 2026. xAI leans into open-ended curiosity and the Grok product line, with a $200 billion valuation stance reported during 2026 funding talks. Thinking Machines sits inside those three, with the collaborative-intelligence framing and a product that supports portability over lock-in. The positioning map across OpenAI, Anthropic, and xAI explains the strong investor demand that the Thinking Machines Lab Series A has attracted.

Comparative strategy also diverges on team shape and governance across these four labs. OpenAI operates with roughly 3,500 employees under a renegotiated corporate structure that places profit alongside its research mission. Anthropic has grown to nearly 1,200 researchers with a Long Term Benefit Trust that oversees mission drift carefully. xAI runs leaner at roughly 500 people, with Elon Musk's personal direction providing both strategic clarity and single-person risk. Thinking Machines is bounded under 300 researchers, with a public benefit corporation structure and a stated commitment to research transparency. The approach matters for enterprise buyers thinking about choosing between AI model providers.

Implementation Lessons From Murati's Second-Act Playbook

Shifting from competitive comparison to action, the Murati second-act playbook offers concrete lessons for founders. First, hire for institutional memory, not just technical range, so your founding team can make calls under crisis conditions on day one. Second, pick a corporate form before the first dollar lands, because changing from a C corp to a public benefit corporation after a round is painful. Third, publish a founding essay that sets measurable guardrails, so investors and recruits know what you will not do, not only what you will do. These three choices compounded for Thinking Machines Lab across the first eighteen months. The playbook is now circulating widely among frontier AI founders who are considering their own exits from established research labs in 2026.

A fourth lesson is sequencing a product launch so it both earns revenue and demonstrates the research philosophy. Tinker's design pulls fine-tuning out of a black box and into the collaborative pattern. That alignment between product and research argument created a self-reinforcing narrative for the Series A fundraising process. Enterprise buyers reported in 2026 that the Tinker purchase decision was tightly coupled to the broader research transparency commitment. Alignment between product and story is now a reproducible playbook element. Founders in adjacent sectors have adopted the same sequencing move.

A fifth lesson concerns public communications discipline for a frontier founder in a crowded market. Murati has kept media appearances rare, with roughly four long-form interviews across the first two years of the company. She has published two signed essays on her lab's blog and co-authored one technical paper. The scarcity makes each appearance count in positioning conversations with regulators, investors, and recruits. The pattern inverts the typical founder flywheel that trades frequent exposure for faster audience growth. The result is a brand with high gravity per publication.

The Risks Baked Into Mira Murati's AGI Bet

Turning from strengths to weaknesses, the risks baked into Thinking Machines Lab are real and worth stating. Concentrated capital at the seed stage creates expectation gravity that can crowd out longer research arcs. The pressure grows as the pathway to artificial general intelligence becomes clearer during the next funding cycle. The lab's dependence on open-weight third-party models for Tinker exposes it to supply-side decisions by Meta, Alibaba, and Mistral teams it does not control. The public benefit corporation structure is novel in frontier AI, with limited case law on how courts will weigh mission against investor duties. These structural risks sit alongside operational ones that the Thinking Machines Lab leadership team tracks on a shared internal dashboard each week. The risk map is well understood internally by the Thinking Machines founding team and shared with the board at each quarterly operating review.

Operational risk includes talent retention at a lab that promised a 300-person cap through 2027 publicly. If frontier research needs scaling beyond that limit, the cap becomes a competitive weakness instead of a cultural asset. There is also the question of whether collaborative general intelligence can match pure autonomous agents on measured tasks that enterprise buyers want to automate. Researchers also point to the policy risk of operating as a smaller lab in jurisdictions that increasingly favor concentration for regulatory simplicity. The broader ethics issues shaking investor confidence across the industry create reputational contagion risk. No single lab can fully insulate itself from category-wide trust shocks.

Ethics, Safety, and Governance in a Post-OpenAI AGI Lab

Building on the risk discussion, ethics and safety are structured into Thinking Machines Lab's design, not bolted on. The lab operates a dedicated safety team of roughly fifteen researchers, led by Lilian Weng. Her reporting line sits parallel to the chief technology officer rather than under product leadership. Pre-deployment red-team exercises run on every major fine-tune released through Tinker and on all internal research model variants. Deployment approvals use a graduated access model, with high-risk capabilities released first to vetted research partners. The approach draws on lessons Murati has described from the DALL-E and ChatGPT deployment cycles at OpenAI. The structure is deliberately not optional for any product launch at the lab and lives inside the organization's formal operating rhythm.

Governance tooling includes a public policy brief that the lab updates each quarter with regulators and civil society organizations. The brief explains the lab's position on model watermarking, synthetic media disclosure, and training data sourcing. External advisors drawn from academia and policy groups participate in quarterly reviews with signed non-disclosure agreements for sensitive capabilities. The structure is more open than OpenAI's current operating model and more formal than most peer startups. It is designed to make the lab easier for regulators to interact with. The transparent AI governance frameworks movement expanded across jurisdictions through 2026.

Internal culture reinforces the governance framework through hiring and performance review. Safety researchers are included on launch review calls at the same seniority as research and product leads. Compensation structures reward safety contributions explicitly, with a dedicated pool for research that identifies risks rather than only ships capabilities. The lab publishes an annual safety report that lists blocked releases, incident response outcomes, and audit findings. That public accounting creates a reputational cost for safety failures proportional to the lab's valuation. The culture translates governance from written policy into daily behavior inside the Thinking Machines Lab engineering and research organizations.

The Future of Mira Murati's AGI Strategy Through 2027

Looking ahead from governance to roadmap, Thinking Machines Lab's 2026 to 2027 strategy centers on three product arcs. Tinker expands from fine-tuning adapters toward full-model customization and multi-model orchestration, with a general availability launch planned for mid-2027. The lab announced a research initiative on long-horizon reasoning that aims to publish benchmarks and open datasets for scientific discovery workflows. A regulated-industry program will package Tinker with audit tooling for healthcare, financial services, and public sector customers. Each arc has a specific revenue, research, and policy outcome. The three product arcs reinforce each other on timing so the lab's revenue, research, and policy outputs land in a coordinated sequence.

Capital strategy is coordinated with the roadmap and the planned $50 billion round. The Series A will carry a longer runway than typical, reportedly 36 months of operating budget at projected spend, so the lab can take research risk without funding anxiety. A planned Series B in 2027 would support a wider enterprise go-to-market team and expanded compute commitments across major cloud providers. Each round carries dilution guardrails to keep research staff ownership above a stated threshold through 2028. The structure resembles a research university endowment more than a conventional startup equity table. The investor base deliberately reinforces that long-horizon posture through governance rights designed to protect research runway from short-term pressure.

Policy strategy focuses on regulated-industry adoption rather than on broad consumer surface area. The lab works with the National Institute of Standards and Technology, the United Kingdom's AI Security Institute, and the European Union's AI Office on evaluation standards. It has committed to public comment periods on major product launches and transparent reporting on known failure modes. The approach positions Thinking Machines as a research-forward lab regulators can engage with directly. That positioning is a bet that regulated adoption will shape the next phase of frontier AI value creation. Mira Murati Leaves OpenAI, Firmly Optimistic About AGI remains the clearest signal of this strategy.

Competitive strategy frames Thinking Machines as a long-horizon counterweight rather than a direct ChatGPT clone. The lab will not release a general-purpose consumer chatbot through 2027, keeping it out of OpenAI's commercial lane. Instead, the Tinker ecosystem and vertical regulated deployments are expected to generate roughly four hundred million dollars of annualized revenue by late 2027. That revenue base, with transparent research and strong governance, is what Murati believes converts her AGI optimism into durable institutional advantage. The next three years will show whether the thesis holds across funding cycles, regulatory shifts, and model progress. The lab is building for a decade, not a product cycle.

Private Valuations of Four Frontier AI Labs, Late 2026

Horizontal bars rank frontier AI labs by latest reported private valuation. Thinking Machines Lab is the newest of the four.

OpenAI$500B
xAI$200B
Anthropic$60B
Thinking Machines Lab target$50B
Thinking Machines Lab seed$12B

Sources: Bloomberg reporting (Thinking Machines target, October 2026), published private-round reporting for OpenAI, xAI, and Anthropic through late 2026. Chart by AIplusInfo.

Key Insights From the Mira Murati OpenAI Transition

The combined signal from these insights is a sharp inversion of leverage between frontier labs and senior researchers. In 2026, a cluster of three or four technical leaders can leave a frontier lab, raise a nine-figure seed within six months, and ship a credible product within a year. That inversion compresses the window in which an incumbent lab can retain scarce institutional knowledge. It also reshapes how investors underwrite early-stage AI valuations around team pedigree rather than product traction. The Murati playbook is now the clearest example of this new founder economics.

Comparing AGI Lab Governance Models After Murati's Exit

Looking across the four most consequential frontier AI labs in late 2026, Thinking Machines Lab holds a distinct governance posture. The comparison table below maps each lab against seven dimensions that enterprise buyers and policy teams track in their diligence conversations. The dimensions include research transparency, external participation, trust mechanisms, decision making, misinformation posture, service delivery, and accountability under corporate structure. Each row translates a public commitment or structural fact into a side-by-side comparison. Enterprise buyers can use the table to shortlist labs that match their regulatory and research requirements. Policy analysts can use it to track governance drift inside the frontier labs through the Series A and Series B rounds.

DimensionThinking Machines LabOpenAIAnthropicxAI
TransparencyQuarterly research digest, annual safety reportSelective capability cards, limited research publicationConstitutional AI papers, model cards for ClaudeOccasional research threads from Elon Musk and leads
ParticipationExternal advisor reviews each quarterRed team contracts, limited external policy inputLong Term Benefit Trust with external trusteesInternal direction with limited external input
Trust mechanismsPublic policy brief updated each quarterMicrosoft deployment partnership, safety committeeResponsible Scaling Policy with capability thresholdsGrok transparency statements, Musk personal signal
Decision makingSafety parallel to CTO, launch review gatesLaunch review with safety committee under CTOPolicy and alignment co-sign launchesFounder-led release decisions
Misinformation postureWatermarking commitment, synthetic media disclosuresC2PA metadata on image and video outputsPublished refusal taxonomy, content provenance workLooser content moderation stance publicly stated
Service deliveryTinker API and vertical regulated-industry programChatGPT platform, API, Microsoft Copilot integrationClaude API, Claude app, enterprise partnershipsGrok consumer app, API, X integration
AccountabilityPublic benefit corporation, annual third-party auditCapped-profit renegotiation to for-profit structurePublic benefit corporation with trust oversightPrivate company under Musk direction

Real-World Examples of Teams Rebuilding After Executive Exits

Three real-world reorganizations show how large AI teams rebuild after a senior executive exit shifts their roadmap. Each example below reports what was implemented, a measurable outcome, a limitation, and a verifiable source for the restructure.

Microsoft AI Division After the Mustafa Suleyman Hire

Microsoft implemented a reorganization of its consumer AI efforts in March 2024, as the CNBC report on the Suleyman Microsoft move describes. The company hired Mustafa Suleyman from Inflection AI to run a new Microsoft AI division. Suleyman arrived with about seventy former Inflection researchers under a $650 million licensing agreement. The reorganization combined Copilot, Bing, and Edge consumer AI products under one leader with authority over roadmap and compute. The measured outcome included a 32 percent year-over-year rise in Copilot active users by Q4 2024 and faster model release cycles. The limitation is that the structure concentrates product authority in one executive hired from a competitor, which creates single-person continuity risk. Microsoft disclosed Copilot usage milestones in its 2024 annual report, giving the public a verifiable track for the restructure's effects.

Google DeepMind Integration Under Demis Hassabis

Google implemented a merger of Google Brain and DeepMind under Demis Hassabis in April 2023 to speed up its frontier AI roadmap. The company announced the move in the Google blog post on the DeepMind consolidation. The merged unit shipped Gemini 1.0 in December 2023 and Gemini 1.5 in February 2024, each delivering a measurable percent increase against standard reasoning benchmarks. By late 2024, the unit had more than 2,500 researchers under one chief and a single roadmap process. The limitation is organizational, because the integration took roughly eighteen months to settle, slowing some mid-tier research arcs during the transition. Google's quarterly disclosures and the Gemini technical reports provide an auditable record of the pre and post integration output. The Gemini release cadence inside the integrated Google DeepMind organization stabilized after mid-2024 with a predictable quarterly delivery pattern on major updates.

Meta FAIR Rebuild After Yann LeCun's Role Shift

Meta implemented a restructure of FAIR in late 2023 as the Meta AI blog post on the FAIR leadership update describes. Yann LeCun's role shifted toward chief AI scientist with less day-to-day operational responsibility. A new generative AI organization under Ahmad Al-Dahle took over Llama development with a 40 percent larger engineering headcount. Measured outcomes included the Llama 3 release in April 2024 and the Llama 3.1 405B model in July 2024, both of which beat prior Meta releases on key benchmarks. The limitation is that FAIR's traditional long-horizon research publication output visibly slowed as researchers moved to product teams. The before and after is observable in Meta's public research page, which still logs publication cadence.

Books that explain the Mira Murati moment

Three verified nonfiction titles that trace the founders, labs, and research race behind her AGI bet.

The Coming Wave: Technology, Power, and the Twenty-first Century's Greatest Dilemma

The Coming Wave: Technology, Power, and the Twenty-first Century's Greatest Dilemma

Mustafa Suleyman's frontier-founder view of the governance questions behind Murati's new lab.

Buy on Amazon
Supremacy: AI, ChatGPT, and the Race That Will Change the World

Supremacy: AI, ChatGPT, and the Race That Will Change the World

Parmy Olson's FT Business Book of the Year 2024 chronicles the DeepMind versus OpenAI race Murati inherited.

Buy on Amazon
Genius Makers: The Mavericks Who Brought AI to Google, Facebook, and the World

Genius Makers: The Mavericks Who Brought AI to Google, Facebook, and the World

Cade Metz's narrative history maps the people and labs behind the September 2024 OpenAI exodus.

Buy on Amazon

As an Amazon Associate, AIplusInfo earns from qualifying purchases.

Case Studies of Research Leaders Who Launched Rival Labs

Three case studies trace research leaders who, like Mira Murati, left established labs to launch rival AGI companies. Each case covers the problem, solution, impact, and limitation that shaped its outcome through late 2026.

Case Study: Dario and Daniela Amodei at Anthropic

The Amodei siblings left OpenAI in late 2020 after disagreements about commercialization pace and safety research priorities, which the Wired profile of Dario Amodei at Anthropic covered in detail. The problem they framed was concentrated commercial pressure crowding out long-horizon alignment work on frontier training runs. They raised a $124 million Series A in May 2021, built a founding team of about twelve former OpenAI colleagues, and framed the lab around constitutional AI safety work. The solution they advanced was a public benefit corporation with a Long Term Benefit Trust that holds a special class of shares to oversee mission drift. By 2024, Anthropic had reached a $18 billion valuation and shipped the Claude model family across API, consumer app, and enterprise channels. The impact includes consistent top-three positioning on independent reasoning benchmarks and a Responsible Scaling Policy framework the industry has adopted in parts.

The limitation is that Anthropic has been pulled into the same compute scaling race it sought to counter. Amazon and Google have each invested more than $8 billion in strategic partnerships. That arrangement creates implicit coupling to hyperscaler infrastructure decisions, which can constrain long-term independence. Anthropic staff have discussed this tension in public panels and the Anthropic's safety-first approach coverage. The company has nonetheless held its research publication cadence and continued to recruit from top universities with strong offers. The Anthropic institutional design, including the Long Term Benefit Trust structure, has survived successive compute shocks across two major funding rounds.

Case Study: Ilya Sutskever at Safe Superintelligence

Ilya Sutskever left OpenAI in May 2024 after the superalignment team disbanded, following the November 2023 board crisis, as a Verge report on the Safe Superintelligence founding explains. The problem he framed was concentrated commercial pressure crowding out long-horizon alignment work on frontier training runs. Safe Superintelligence launched in June 2024 with a stated single-product focus and a public commitment to no commercial product release before alignment research matured. The lab raised more than $1 billion by mid-2024 and reached a $32 billion valuation by late 2025 on a disciplined headcount of under 25 researchers. The impact is a counter-model to Thinking Machines, one that chooses research-only operations over product distribution.

The limitation is that Safe Superintelligence has not shipped anything auditable to the research community through late 2026, which raises questions about verifiable progress. Critics argue that the lab's secrecy inverts the open publishing culture that produced the transformer and the scaling-laws era of research. Sutskever has stated that the lab will publish a safety report before any model is shared publicly, with a commitment he has repeated in rare appearances. The team has nonetheless hired aggressively from Google DeepMind and OpenAI, suggesting researcher confidence in the approach. The gamble on secrecy will be judged by its first public artifact.

Case Study: Noam Shazeer at Character AI and the Google Return

Noam Shazeer co-founded Character AI in 2021 after leaving Google, framing the problem as low-latency conversational AI that mainstream labs were ignoring at scale. Character AI reached roughly 20 million monthly active users and a $1 billion valuation by 2023, as the TechCrunch report on the Character AI valuation documented. The solution was a product-first approach to persona-rich chatbots that let users shape model behavior through conversation rather than fine-tuning. In August 2024, Google announced a $2.7 billion licensing deal that returned Shazeer to Google and gave Character AI a durable commercial path. The impact is a measured return of a prominent researcher to a hyperscaler, with a technology transfer deal that preserves the startup's product.

The limitation and controversy is the signal the deal sent about startup independence under extreme talent demand. Policy groups argued that the licensing structure allowed Google to absorb a competitor's research team without a merger review. The Reuters report on the FTC inquiry into AI partnerships documented that broader regulatory posture at the time. The arrangement nonetheless gave Shazeer a research platform at Google's scale and gave Character AI runway to pursue its product thesis. The case study shows a different exit pattern from the Murati playbook, with corporate absorption rather than independent scaling. Both the independent-scaling and corporate-return patterns remain available to frontier researchers in 2026 depending on their appetite for standalone operation.

Frequently Asked Questions About Mira Murati's AGI Bet

Why did Mira Murati leave OpenAI?

Mira Murati left OpenAI in September 2024 to pursue her own research exploration outside the company. She framed the decision as a personal reset after six years leading product and research. She wanted room to build an organization around her collaborative general intelligence thesis.

What is Thinking Machines Lab?

Thinking Machines Lab is Mira Murati's AI research company, founded in February 2025 in San Francisco. It builds collaborative general intelligence tools and shipped Tinker in October 2025. The lab operates as a Delaware public benefit corporation with a stated 300-researcher cap.

How much did Thinking Machines Lab raise in its seed round?

Thinking Machines Lab closed a $2 billion seed round in July 2025 at a $12 billion valuation. Andreessen Horowitz led the round, with Nvidia, Accel, and ServiceNow participating. It was the largest seed round in startup history at the time.

What is Tinker from Thinking Machines Lab?

Tinker is a hosted fine-tuning API that lets teams customize open-weight frontier models with minimal infrastructure work. It supports Llama, Qwen, and Mistral model families at launch. An open source cookbook on GitHub bundles reference implementations for common fine-tuning tasks.

Who is on the Thinking Machines Lab founding team?

The founding team includes John Schulman as chief scientist and Barret Zoph as chief technology officer. Andrew Tulloch, Jonathan Lachman, and Lilian Weng filled out senior leadership. All five had worked together at OpenAI on major 2022 to 2024 projects.

What is collaborative general intelligence?

Collaborative general intelligence is an architectural stance in which a human expert and a trained model exchange structured artifacts across a reasoning session. The model flags uncertainty, defers where appropriate, and surfaces reasoning paths a reviewer can audit. The pattern changes evaluation suites, training objectives, and interface design across the full research and deployment lifecycle.

How does Thinking Machines Lab compare to OpenAI?

Thinking Machines Lab is smaller, more research transparent, and focused on collaborative intelligence rather than consumer scale. It will not release a general-purpose consumer chatbot through 2027. The lab's $12 billion seed valuation is roughly one-fortieth of OpenAI's late 2026 private valuation.

Why is Mira Murati optimistic about AGI?

She sees AGI as a joint capability humans and models build together, not a system one lab announces finished. Her view is that open publishing, strong safety ownership, and shared equity keep AGI aligned with public interest. The stance shapes how her lab organizes research and governance.

What did Mira Murati do during the November 2023 OpenAI board crisis?

Mira Murati served as OpenAI interim CEO for roughly 72 hours after the board fired Sam Altman on November 17, 2023. She coordinated with the board, Microsoft, and 700 employees who signed a letter threatening to resign. She then stepped back to the CTO role once Altman agreed to return under a reconstituted board.

What was the September 2024 executive exodus at OpenAI?

Three senior OpenAI leaders resigned within one week in September 2024. Mira Murati, chief research officer Bob McGrew, and vice president of research Barret Zoph all announced departures. The cluster followed capped-profit renegotiation and the disbanding of the superalignment team.

Will Thinking Machines Lab compete with ChatGPT?

Thinking Machines Lab will not release a general-purpose consumer chatbot through 2027, keeping it out of ChatGPT's commercial lane. The product strategy focuses on the Tinker platform and regulated-industry adoption rather than consumer chatbot distribution. Enterprise customization and vertical deployments are its primary revenue arcs.

How does the Thinking Machines safety team work?

The safety team of about fifteen researchers reports in parallel to the CTO, not under product. Pre-deployment red-team exercises run on every major fine-tune and research model variant. Deployment approvals use a graduated access model with vetted research partners first.

What is the Thinking Machines Lab Series A valuation target?

Bloomberg reported in October 2026 that Thinking Machines Lab was raising at a $50 billion target. The round would place the company inside the top five private AI valuations in the world. Round participants include new strategic partners alongside expanded compute commitments across major cloud providers.

How can I try Tinker from Thinking Machines Lab?

Tinker offers a hosted fine-tuning API accessed through a Python SDK, with a free tier for research users. The open source cookbook on GitHub explains supervised fine-tuning, direct preference optimization, and reward-model-free alignment. Pricing starts at a flat per-token adaptation fee for standard usage, with volume breaks for recurring experimentation runs.