Reid Hoffman and Mark Pincus Back Computer-Use AI Lab Chasing $1 Billion Valuation
Prentis is training models to automate office workflows by watching how workers navigate documents and systems, with early customers already signed for millions in contracts.
A New Bet on Automation
Ritankar Das spent his twenties building and selling AI companies through Titan, a holding company he founded after abandoning a Cambridge PhD. Now 31, he is leading Prentis, an AI research lab that wants to replace the tedious parts of office work with software that can actually operate computers. The startup, launched in April with backing from LinkedIn co-founder Reid Hoffman and Zynga founder Mark Pincus, is currently negotiating a $100 million funding round that would value the company at $1 billion, according to people familiar with the discussions.
Unlike chatbots that generate text or summarize documents, Prentis is focused on computer-use models: AI systems trained to watch how office workers click through applications, fill forms, and hunt for information across different platforms. The goal is to build agents capable of handling end-to-end workflows without human supervision. Insurance claims processing, customs duty exceptions, and document-heavy compliance tasks are among the early use cases the company is targeting.
Early Traction in Enterprise
Prentis has already secured contracts with several customers, including a healthcare management service organization, a manufacturer, and clothing producers. The combined value of these agreements reaches up to $50 million, according to sources. Internal projections suggest the company could hit a $75 million annualized run rate by the third quarter of this year, though those figures are tied to performance-based pricing: Prentis charges customers roughly 20 percent of the savings its agents deliver, meaning revenue depends on execution rather than upfront licensing fees.
That pricing model reflects both confidence and pragmatism. If the agents fail to deliver measurable efficiency gains, Prentis does not get paid. If they succeed, the startup captures a slice of the cost reduction its software creates. For customers wary of AI hype, the structure offers a lower-risk entry point.
Competing on Cost, Not Scale
Prentis claims its Hive-32B model outperforms larger frontier systems on two computer-use benchmarks: WindowsAgentArena, which tests task completion in real Windows applications, and ScreenSpot-v2, which measures an agent's ability to identify the correct on-screen control. According to the company, Hive-32B beats OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on these tests while running at roughly one-tenth the cost per task.
The economic argument matters as much as the technical one. Frontier models from OpenAI and Anthropic are expensive to run at scale, particularly for repetitive, high-volume tasks like processing insurance paperwork or reconciling invoices. Prentis is betting that a smaller, more efficient model optimized for computer control can deliver comparable accuracy at a price point that makes automation viable for everyday workflows, not just high-value use cases.
We have not independently verified the benchmark claims, and the company declined to provide additional data or commentary.
A Crowded Field with Deep-Pocketed Rivals
Prentis is entering a market where competition is intensifying. Anthropic, OpenAI, and Mira Murati's Thinking Machines Lab are all developing AI agents designed to operate computers autonomously. Earlier this year, Anthropic acquired Vercept, a Seattle-based startup focused on computer-use models, folding its team into its own research efforts and shutting down the product. The acquisition signals that larger labs see computer control as a strategic capability worth building or buying aggressively.
The challenge for Prentis is differentiation. Anthropic and OpenAI have vast compute resources, established enterprise relationships, and brand recognition. What Prentis offers is focus: a company built exclusively around computer-use models rather than a general-purpose AI lab exploring multiple research directions. That specialization could translate into faster iteration and deeper domain expertise, but it also means the startup has less room for error.
The Founders Behind the Lab
Das brings an unusual profile to the CEO role. He graduated from UC Berkeley at 18 with a double major in bioengineering and chemical biology, earning recognition as the university's youngest medalist in over a century. After completing a master's in biomedical engineering at Oxford, he started a PhD in AI at Cambridge as a Gates Cambridge Scholar before dropping out in 2014 to launch Titan.
Titan operates as a holding company for AI startups, funded by exits rather than external limited partners. The model echoes Berkshire Hathaway's structure, with Titan building, operating, and occasionally selling companies rather than relying on venture capital. Previous Titan ventures include Tala Health, an AI-powered virtual care provider that raised $100 million, and Forta Health, an autism care startup that raised $55 million. Dascena, a disease prediction company launched under Titan, was acquired by CirrusDx in 2022.
For Hoffman and Pincus, Prentis is a side project. Hoffman recently stepped down from Microsoft's board after nearly a decade to focus on Manas AI, a drug-discovery startup he is backing and helping to build. He was an early OpenAI investor and co-founded Inflection AI with Mustafa Suleyman before Microsoft absorbed most of that team in 2024. Pincus now runs Reinvent Capital, an investment firm where Hoffman serves as senior adviser. Both bring networks, capital, and credibility, but the day-to-day execution falls to Das and the team he is assembling.
Building the Team
Prentis has hired more than 25 employees, drawing researchers from OpenAI, Google DeepMind, Meta, Tencent, and Alibaba. The talent concentration suggests the company is serious about competing on research quality, not just product velocity. In a field where model performance can shift rapidly with new architectures or training techniques, having engineers who have built state-of-the-art systems elsewhere is a meaningful advantage.
The Automation Thesis
Prentis is making a bet that automating office work will become a larger AI market than coding assistance. That thesis is not universally shared. Coding remains one of the most successful early applications of large language models, with tools like GitHub Copilot and Cursor seeing widespread adoption among developers. Office automation, by contrast, requires models to navigate messy, unstructured environments where applications, data formats, and workflows vary widely across companies.
If Prentis is right, the market opportunity is enormous. Knowledge workers spend significant time on repetitive tasks that software could plausibly handle: copying data between systems, generating routine reports, tracking down documents, and chasing approvals. The question is whether computer-use models are reliable enough to handle these tasks autonomously, or whether they will require constant human oversight that negates the efficiency gains.
The answer will depend on execution. A model that completes 95 percent of a task correctly but fails unpredictably on the remaining 5 percent may not be much more useful than no automation at all. Prentis will need to prove not just that its agents can perform well on benchmarks, but that they can operate reliably in production environments where mistakes have real consequences.
The $1 billion valuation reflects optimism that Das and his backers can deliver on that promise. Whether the early customer contracts translate into sustained revenue growth will determine if that optimism was justified.
