Prentis seeks $100M at $1B valuation for computer-use agents

Why it matters

Prentis is testing whether smaller models trained inside real workflows can beat frontier systems on cost. Its valuation pitch depends on performance-linked contract value, not booked revenue.

Ritankar Das, Reid Hoffman (@reidhoffman) and Mark Pincus are seeking $100 million for Prentis, an AI research lab built around a bet that routine computer work will become a larger market for AI agents than software coding.

Prentis was in funding talks at a proposed $1 billion valuation as of July 24th, TechCrunch reported, citing two people familiar with the discussions. The financing has not closed, and no lead investor or participating firms were named.

The proposed price would give Prentis unicorn status roughly three months after its April launch. That speed rests on an aggressive commercial case: TechCrunch reported that Prentis has signed contracts worth up to $50 million with several unnamed customers across healthcare management, manufacturing and apparel. Investor materials reviewed by TechCrunch project an estimated $75 million annualized run rate by the third quarter of 2026.

Those figures require a narrower reading than a conventional revenue forecast. Prentis would receive fees equal to 20% of the savings generated for customers, according to the pitch deck. The $75 million estimate represents annualized contract value tied to expected performance, not revenue Prentis has already recognized. The materials describe the estimate as dependent on performance and final execution.

Das builds another operating company

Das is the operating founder and Prentis CEO, while Hoffman and Pincus bring two of Silicon Valley's most recognizable founder and investor networks.

Das has spent much of his career forming AI businesses around specific industries. He founded Titan Holdings, which describes itself as a holding company that builds and operates AI-native companies. Titan says it has five active operating companies, over 300 employees across its portfolio, two exits and two companies valued above $1 billion. Those are Titan's own figures.

His academic career moved just as quickly. UC Berkeley named Das its University Medalist in 2013 after he completed double majors in bioengineering and chemical biology, plus a creative-writing minor, at 18. Berkeley said he graduated in three years with a 3.99 GPA and became its youngest University Medalist in at least a century. He later earned a master's degree in biomedical engineering at Oxford and left a Gates Cambridge PhD program to build companies, according to TechCrunch.

Titan's earlier businesses provide the clearest guide to Das's approach. Forta Health announced a $55 million Series A, led by Insight Partners, in 2024. CirrusDx acquired Dascena's Texas laboratory business in 2022. Titan also lists Tala Health among its operating companies.

Prentis extends that company-building model into horizontal workplace automation. Das is pursuing workflows shared across large organizations, while starting with operational problems that require specialized knowledge: insurance claims, customs-duty refund exceptions and processes spread across documents and multiple software systems.

Hoffman and Pincus give that effort institutional reach. Hoffman co-founded LinkedIn, joined Greylock and became an early OpenAI backer before co-founding Inflection AI. RuntimeWire reported in June that Hoffman would leave Microsoft's board, nearly a decade after Microsoft acquired LinkedIn. Pincus founded Zynga and later formed Reinvent Capital.

Training agents inside the work

Prentis's thesis starts with a practical problem: the instructions that define office work are incomplete. Exceptions, handoffs and recovery steps often live in employee habits rather than manuals or public training data.

Prentis says it trains computer-use models inside those workflows. The models perceive interfaces, take actions across mobile, browser and desktop software, and learn from results, including failed attempts. Prentis lists over 25 employees and says its researchers have worked at OpenAI, Google DeepMind, Meta, Tencent and Alibaba.

That approach is aimed at tasks where connecting through an API is difficult or insufficient. A computer-use agent can work through the same interface as an employee, clicking controls, reading documents and transferring information between systems. It also inherits the risks of that interface: an incorrect click or misunderstood exception can alter a record, submit a form or send the workflow down the wrong path.

Prentis calls its model Hive-32B. According to the pitch deck, Prentis claims Hive-32B beats OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on WindowsAgentArena and ScreenSpot-v2, benchmarks that test task completion and on-screen control selection. Prentis also claims approximately one-tenth the cost per task of frontier model APIs. TechCrunch did not independently verify either claim, and Prentis has not published enough supporting material for outside researchers to reproduce the comparison.

The smaller-model strategy matters because enterprise automation is sensitive to unit economics. An agent that saves labor on a single task can still be uneconomic if it repeatedly calls a costly frontier model, requires frequent human intervention or fails often enough to create cleanup work. Prentis is pitching a tighter loop: train on the customer's actual environment, specialize the model and charge against measured savings.

Computer use is already a platform fight

Prentis is entering a category where model developers and cloud providers are buying teams and shipping their own automation stacks. Anthropic introduced computer use for Claude in 2024, allowing the model to inspect screens, move a cursor and type. In February 2026, Anthropic acquired Vercept, a Seattle computer-use specialist, and folded its founders into Anthropic while winding down Vercept's external product.

Amazon has made Nova Act generally available for browser-based production workflows, while Google has introduced a Gemini computer-use model. Startups including Simular and H Company are also selling agents designed to operate graphical interfaces.

Prentis is therefore raising against well-funded platforms that control models, cloud distribution and enterprise relationships. Das's answer is specialization: collect experience from real workflows, run a smaller model and tie Prentis's economics to the savings produced.

The funding talks test whether investors will price that model on contracted potential before Prentis establishes recognized revenue, discloses customers or publishes reproducible technical results. The proposed $1 billion valuation says as much about the appetite for computer-use agents as it does about Prentis's first three months. Closing the round would give Das the capital to turn a set of performance-linked contracts into proof that AI can reliably operate the software where office work already happens.