Prentis Hive-32B Tops GPT-5.4 at 10x Lower Cost on PC Benchmarks

In the escalating war to build AI agents that actually use computers, a new lab claims a dual win: its 32-billion-parameter model not only tops the leaderboards but does so at a fraction of the cost. Prentis says its Hive-32B beats GPT-5.4 and Claude Opus 4.6 on WindowsAgentArena and ScreenSpot-v2, potentially shifting the economics of enterprise AI deployment toward smaller, specialized models.

Prentis, a new AI research lab co-founded by LinkedIn’s Reid Hoffman, Zynga’s Mark Pincus, and serial entrepreneur Ritankar Das, is in advanced discussions to raise $100 million at a $1 billion valuation, according to two people familiar with the talks. The funding round, which would value the company at unicorn status just three months after its April 2026 launch, underscores the intense investor appetite for AI agents capable of controlling computers to automate enterprise workflows. Prentis is training models that observe and replicate how office workers navigate documents, applications, and systems, with the goal of delivering AI agents tailored to specific industry tasks—such as processing insurance claims or handling customs duty refunds.

“

From a market perspective, the $100 million raise at a $1 billion valuation only months after launch signals a continued frothy environment for top-tier AI startups, even as some sectors of venture capital cool.

The startup has already demonstrated early commercial traction, having signed contracts worth up to $50 million with customers in healthcare management, manufacturing, and goods/clothing sectors. These contracts, according to investor materials obtained by TechCrunch, are structured such that Prentis earns a fee equal to 20% of the savings its agents generate for clients. While the company projects an estimated $75 million annualized run rate by the third quarter of 2026, it cautions that these figures are “performance-dependent and subject to final execution,” meaning realized revenue may materially differ from the forecast. This performance-contingent model aligns incentives but also introduces execution risk; if the AI agents fail to deliver savings, Prentis’ revenue would be directly impacted.

The lab’s technical strategy centers on a smaller, more efficient model—Hive-32B—which it claims outperforms leading frontier systems on two 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 locate the correct on-screen controls. According to Prentis’ pitch deck, Hive-32B not only beats OpenAI’s GPT-5.4 and Anthropic’s Claude Opus 4.6 on these benchmarks but does so at roughly one-tenth the cost per task. This cost efficiency could be a significant competitive moat, as many enterprises are sensitive to the per-transaction economics of deploying AI agents at scale. By focusing on a specialized computer-use domain and training a leaner model, Prentis aims to undercut general-purpose systems that are often overkill for routine office tasks.

The involvement of Hoffman and Pincus brings not only star power but also deep operational and network advantages. Hoffman, a co-founder of LinkedIn and a prolific venture capitalist at Greylock, has long bet on AI-driven transformation of work. Pincus, who created the social gaming giant Zynga, brings experience in scaling consumer-centric platforms that can be adapted to enterprise interfaces. Their combined credibility is likely accelerating investor interest and customer adoption. However, the AI agent space is crowded, with OpenAI, Anthropic, Adept, and others pouring resources into computer-use capabilities. Prentis’ ability to defend its niche will depend on execution speed, the defensibility of its training data (derived from observing real workflows), and the sticky integration with client systems.