Prentis — New AI Lab in Talks to Raise $100M

Reid Hoffman and Marc Pincus — the LinkedIn and Zynga co-founders who built two of the defining internet platforms of the 2000s — have quietly backed a new AI lab called Prentis, which TechCrunch reported on July 24, 2026, is in talks to raise $100 million at a $1 billion valuation. The company, launched in April under 31-year-old CEO Ritankar Das, is training a 32-billion-parameter model called Hive-32B to navigate real office software — not write code or generate prose, but actually click buttons, fill forms, and complete workflows the way a human employee would. If its benchmark claims hold under independent scrutiny, Prentis represents the most commercially direct bet yet on AI that operates computers rather than merely advises the people who do.

What's Converging

The computer-use category has moved faster in the last eight months than most practitioners expected. Anthropic shipped its computer-use API in late 2024, then followed it up by acquiring Vercept, a Seattle startup focused specifically on GUI-operating agents — and subsequently shut Vercept down entirely, folding the team inward. That move signals something concrete: Anthropic no longer views computer use as a third-party integration problem. It's treating it as core infrastructure worth absorbing. OpenAI has been running parallel tracks, embedding screen-navigation capabilities directly into its GPT-5-era models. Mira Murati's Thinking Machines, according to reporting across multiple outlets covering Prentis's launch, is also building in this direction.

What all of these efforts share is a common premise: the next major AI use case isn't another chatbot or coding assistant — it's software that removes entire job functions from the human-in-the-loop. Office workers spend an outsized share of their time doing navigational work across the same applications, week after week. Claims processing, customs documentation, purchase order reconciliation: these are tasks rigid enough in structure that a well-trained model should be able to complete them without human oversight at every step. The commercial pull is not subtle.

What marks this particular moment as distinct from earlier computer-use announcements is the profile of the investors now placing real money behind the thesis. Hoffman has historically been an early-money signal for transformational network-effect businesses. His involvement alongside Pincus doesn't guarantee Prentis succeeds, but it does confirm this category has cleared the "serious enough to back" bar for some of the most pattern-experienced investors in the industry.

The Specific Development

Prentis launched in April 2026 with Ritankar Das — described in TechCrunch's reporting as the youngest University Medalist in UC Berkeley's history in over a century, and the founder of a self-funded holding company modeled loosely on Berkshire Hathaway's structure — alongside Hoffman and Pincus as co-founders. The company's central technical offering is Hive-32B, a 32-billion-parameter model trained specifically on how office workers navigate real workflows across documents, systems, and interfaces. Startup Fortune's coverage adds that the model is designed to operate across mobile, desktop, and browser environments, not just Windows. The pitch is precision over scale: rather than applying a 500-billion-parameter frontier model to an insurance claims queue and hoping for the best, train a smaller model to be reliably good at exactly that task.

The financial picture confirmed across multiple reports — including TechCrunch's primary account, corroborated by Startup Fortune and TickrWire — shows Prentis has signed contracts worth up to $50 million with early customers in healthcare management, manufacturing, and goods-and-clothing. The company's investor materials project $75 million in annualized run rate by the third quarter of 2026. Critically, those figures are performance-dependent projections, not recognized revenue: Prentis's pitch deck states that contracted fees equal 20% of savings realized, meaning customers pay based on what the system actually delivers. That structure is either a compelling alignment of incentives or a meaningful revenue risk, depending on how precisely Prentis can control outcomes across varied enterprise deployments.

On the technical side, Prentis claims Hive-32B outperforms both OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on two specific computer-use benchmarks: WindowsAgentArena, which measures end-to-end task completion on real Windows applications, and ScreenSpot-v2, which tests a model's ability to identify and activate the correct on-screen element. The company also claims costs approximately 10 times lower per task than frontier API pricing. TechCrunch explicitly noted it has not independently verified any of these results — a caveat that should stay front of mind before treating Hive-32B's reported performance as settled fact.

Our read is that the cost argument is the more structurally durable of the two claims, regardless of where benchmark rankings ultimately land. Frontier API pricing for high-volume agentic workflows is genuinely prohibitive at scale. A model that costs one-tenth as much per task and completes that task with adequate reliability is often more commercially viable than a more capable but far more expensive alternative — particularly when the task is narrow enough that "adequate" is achievable with a focused training approach. The benchmark competition with GPT-5.4 and Opus 4.6 generates the headlines. The unit economics will determine whether the $75 million ARR projection survives contact with actual enterprise procurement cycles.

What's Likely Next

The most immediate question is whether the $100 million round closes at the reported $1 billion valuation and who leads it. Prentis has not disclosed lead investors, and the "in talks" framing in all reporting means nothing is signed. Given the competitive landscape, investors will be assessing whether Prentis's moat lives in the model architecture itself, the domain-specific training data it has accumulated, its early enterprise relationships, or Das's stated plan to run a holding-company model of targeted acquisitions in adjacent workflows. Each of those would be defensible, but they point to different valuations and different diligence timelines.

In the 30-to-90-day window, the more revealing signal will be whether independent evaluators or early enterprise customers can confirm the WindowsAgentArena and ScreenSpot-v2 results in production conditions rather than curated test sets. Vendor-reported benchmarks are a starting point. More pressingly, the performance-dependent revenue model means that if Hive-32B encounters friction on first deployments, that will surface in renegotiations or customer disclosures before the end of Q3 2026 — precisely when the projected $75 million annualized run rate is supposed to materialize. That timeline is tight, and it makes the next ninety days a genuine stress test of whether Prentis's commercial thesis holds outside the pitch deck.

Sources