Former OpenAI and Anthropic employees are forming a new generation of AI startups centered on verification, safety, automation and hardware, placing less emphasis on building another frontier large language model from scratch. The pattern is emerging as the two leading AI labs prepare for potential initial public offerings and as former staff gain liquidity from large private share sales.
An OpenAI employee share sale in autumn 2025 allowed staff to sell a combined $6.6 billion of stock, valuing the company at $500 billion. That event has given early employees more freedom to finance new ventures or leave established roles without relying solely on traditional venture backing. Anthropic, itself founded roughly five years ago by former OpenAI employees, has reached a reported $380 billion valuation, underscoring the scale of capital now surrounding the leading AI labs.
Silicon Valley has compared the movement to the PayPal alumni network that emerged after eBay acquired the payments company for $1.5 billion in 2002. The comparison fits in one respect: a concentrated group of engineers, researchers and operators has accumulated technical experience and financial resources inside a few high-value companies, then begun applying those resources to new businesses.
Startups target the layers around AI models
Many of the new companies are focused on making AI systems more useful, dependable or affordable to deploy. That reflects a practical constraint facing the industry: producing a response from a powerful model can be relatively simple, while proving the response is correct, connecting it to business tools or preventing failures remains difficult.
Tworek, a former OpenAI research vice president, founded Core Automation to develop an automated research laboratory. The company’s proposed system would use models to read scientific papers, generate hypotheses and run experiments. Such an approach aims to turn AI from a research assistant into a more active component of the scientific process, though it would depend on reliable experimental design and validation rather than language capability alone.
Neyshabur, formerly a researcher at Anthropic, helped establish Mirendil, which has raised $200 million. Mirendil is developing systems intended to speed up improvements in AI models while reducing the amount of direct human research work required. The funding size shows that venture firms are willing to place substantial bets on teams that could improve the process of AI development itself, rather than only creating applications on top of existing models.
The emphasis on automated research also carries a commercial attraction. A startup that can make model evaluation, experimentation or data generation less labor-intensive could sell to AI labs, enterprise software groups and research organizations without having to compete directly with OpenAI, Anthropic, Google or other companies spending heavily on foundation-model training.
Verification has become a costly AI bottleneck
One cluster of former lab employees is concentrating on verification, the process of determining whether an AI system’s output is accurate. In many use cases, checking an answer can take more time and expertise than generating it, particularly in mathematics, software engineering, science and legal or financial work.
Han, a former OpenAI researcher, founded Math Inc to convert mathematical proofs into forms that machines can check line by line. Formal proof systems have long been used in specialized mathematical and software-verification settings, but AI has increased interest in applying them more broadly. A model may produce a plausible-looking proof, while a formal verifier can test whether every step follows the permitted rules.
That distinction could become increasingly valuable as companies try to use AI in work where an incorrect answer carries a high cost. A business may accept an imperfect chatbot response, but cannot easily tolerate flawed calculations in engineering, research or security-sensitive code. Startups that build checking systems could become part of the infrastructure required before autonomous AI tools are trusted with more consequential tasks.
Agent builders focus on execution and personal use
Other alumni-led companies are building AI agents, systems designed to break down tasks, use external tools, retain context and complete multi-step workflows. Rational and Zavify, founded by former OpenAI and Anthropic employees, are pursuing this operational layer of AI deployment.
The technical challenge is broader than selecting a model. An agent used in a workplace may need access to internal data, scheduling systems, documents, software tools and human approval processes. It must also recover from errors and preserve a clear record of what it did. Those requirements create opportunities for startups that can package models into reliable workflows rather than simply offering another chat interface.
Babuschkin, a former xAI co-founder, has launched River AI, a personal AI product intended to adapt to individual users. Consumer AI products face a different test: they need to earn enough user trust to handle private context while providing clear daily utility. Personalized systems could become more useful as they retain preferences and habits, though privacy, data handling and user control will shape whether such products gain traction.
Hardware has also drawn former AI-lab talent. Edrisian, previously an OpenAI Codex engineer, launched Blackstar to develop a personal computer designed around AI-era use cases. The effort suggests that some founders see limitations in today’s general-purpose devices, especially for software that uses local models, persistent assistants or AI-driven interfaces.
Safety startups move from internal teams to the market
A separate group of companies is emerging from safety and alignment work previously conducted inside major labs. Syntony, formed by a former Anthropic team, is designed to test models through adversarial probing for failures and boundary breaches.
Resolution, founded by a former OpenAI researcher, is working on methods to assess whether a model is acting in line with human intent and to assign confidence scores to its actions or answers. Guidelight, also founded by former OpenAI employees, is developing safety standards and compliance approaches for the industry.
These companies are entering a market where AI developers face growing pressure from enterprise customers and regulators to explain how systems are tested, monitored and controlled. Safety practices that once sat mainly inside research labs could become independent products and services, particularly for companies deploying third-party models in regulated or high-risk environments.
Liquidity and fundraising accelerate the breakouts
Large fundraising rounds are helping the alumni companies move quickly. Mirendil’s $200 million round and the reported $2 billion raised by Murati, despite no product being disclosed in the supplied information, illustrate how strongly venture capital is following experienced AI-lab teams.
Venture firms are also monitoring departures closely. Rosenthal, OpenAI’s first head of sales, has moved into investing, while Deng, a former consumer product lead, has joined Felicis. Their moves reflect a growing market for people who understand how major AI labs recruit talent, build products and identify technical gaps.
The wave of spinouts does not establish that frontier model development has become irrelevant. Training leading models remains capital-intensive and dominated by a small number of well-funded companies. Yet the new startups show that much of the commercial opportunity is moving into the systems that evaluate models, give them access to tools, make them safer and place them into daily work.
Explore how emerging AI tools are reshaping crypto trading—start with our guide on AI copy trading and automation opportunities.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
