Google Closes $1.5 Billion Mechanize Deal as Startup's Cofounder Joins DeepMind

Google has closed its $1.5 billion Mechanize deal, bringing cofounder Tamay Besiroglu and a dozen-plus staffers into DeepMind to work on midtraining and coding.

Google has completed its deal for Mechanize, the startup that builds simulated environments for training AI agents, bringing cofounder and former CEO Tamay Besiroglu and more than a dozen of the company's staff into Google DeepMind, according to Business Insider. The agreement, valued at roughly $1.5 billion or more, is structured as a talent hire plus a technology license rather than an acquisition — a deal type that has become Google's template for absorbing promising AI teams without triggering a regulatory review.

The completion closes a chapter that opened in early August, when Business Insider first reported that Google was in talks for a $1.5 billion-plus arrangement with the San Francisco-based startup. The new reporting, published Friday, confirms the deal has effectively closed, with the staffing moves visible on the LinkedIn profiles of former Mechanize employees. It marks one of the most consequential talent transactions of the year in the race to build agents that can do real computer work, a competition covered daily on AI Buzz Wire.

Who is moving where

Besiroglu, who co-founded Mechanize and served as its chief executive, is now a research scientist at Google DeepMind, according to his LinkedIn profile. He is joined by more than a dozen former Mechanize staffers, the majority of whom have joined Google's midtraining efforts — the increasingly competitive discipline of shaping model behavior between pretraining and product release, where much of the recent progress in coding and agentic capabilities has concentrated.

Back at the startup, Guive Assadi, who served as Mechanize's chief of staff, has taken over as CEO. That structure — the acquirer absorbing most of the technical team while a small leadership remnant keeps the company entity alive — mirrors Google's earlier arrangements with Character.AI and the coding startup Windsurf, both of which were designed to secure talent and technology without a formal merger or acquisition.

What Mechanize actually builds

Mechanize's core product is the training environment itself. The company constructs what researchers describe as "digital offices" — simulated software environments in which an AI agent is given a realistic workplace task, such as navigating an unfamiliar application or completing a multi-step workflow, and receives a reward signal based on whether it succeeds.

The approach reflects a thesis that has gained wide currency across the industry: the binding constraint on agent capabilities is not raw model scale but the availability of rich, verifiable environments in which agents can practice. TechCrunch has described the work as building "training grounds that simulate what an AI agent would be doing in a real software application," comparable to "creating a very boring video game." Reinforcement learning on such environments has been credited with many of the recent gains in coding benchmarks and long-horizon task completion.

The environment-building niche has quietly become one of the most contested supply chains in AI. On the Dwarkesh Podcast this week, discussion of how quickly firms can scale up the complexity of the reinforcement learning environments they train on — Mercor was among the companies named — framed those teams as a genuine bottleneck on frontier progress. Google paying nine figures' worth of attention to a startup with a headcount in the dozens underlines how scarce the expertise has become.

A structure built for the antitrust era

The non-exclusive technology license at the heart of the deal means Mechanize continues to exist as an independent company and can, in principle, work with other customers. That detail matters: non-exclusivity is a hallmark of the deal structures Google and other tech giants have favored since regulators began scrutinizing conventional AI acquisitions more aggressively. Microsoft's arrangement with Inflection and Google's own Character.AI and Windsurf deals all followed the same logic — pay billions for the people and the technology, avoid the word "acquisition."

Critics have argued such deals are designed to sidestep antitrust review, while defenders note that the talent genuinely moves and the startups often continue operating. Either way, the Mechanize agreement is among the largest of this deal type to reach completion, and its close suggests the structure has survived contact with the current regulatory environment.

Why coding is the battleground

Google's stated motivation, according to the August reporting, was improving the coding capabilities of its Gemini models — an area where the competition with OpenAI and Anthropic has been measured in benchmark decimal points and enterprise contract wins. Automated AI research is the natural endpoint of that race: the Dwarkesh Podcast episode this week featured OpenAI co-founder John Schulman estimating that recursive self-improvement could arrive within three to four years, predicated on exactly the kind of practice environments Mechanize specializes in building.

For Google DeepMind, the acquisition of a proven environment-building team is a bet that the next capability jump will come from the training loop rather than the base model. For the rest of the industry, the deal's completion sets a price anchor for what a top-tier reinforcement learning environment startup is worth — and a signal that the talent land grab of 2026 is far from over.