Google has absorbed Mechanize’s coding-eval team into DeepMind after $1.5B+ deal talks

Google appears to have completed a talent-focused deal with Mechanize, the startup that builds training environments and evaluations for frontier coding agents. Business Insider reports that Mechanize co-founder and former CEO Tamay Besiroglu is now a research scientist at Google DeepMind and that more than a dozen former Mechanize employees have joined Google, with many working on model midtraining. The final financial terms have not been disclosed. Earlier reporting said Google was discussing a package worth more than $1.5 billion for Mechanize technology and talent. Mechanize itself still has an active website and hiring pages, and its former chief of staff Guive Assadi is now described as CEO in public profiles, so this is not a conventional whole-company acquisition.

Mechanize is not primarily a consumer coding assistant. Its product is the harder-to-see layer behind coding models: environments, long-horizon software-engineering tasks, graders and reinforcement-learning signals. Moving much of that team into Google suggests that coding-agent progress is increasingly constrained by the quality of training tasks and evaluations, not just model scale. For small AI infrastructure companies, the reported economics are also striking: the market can value a compact team that knows how to make frontier coding systems measurably better far above normal SaaS multiples, even when the buyer does not acquire the entire company.

Business Insider reports that Tamay Besiroglu has joined Google DeepMind as a research scientist and that more than a dozen former Mechanize employees have moved to Google, many into midtraining work. Google and Besiroglu declined to comment on the final deal terms. Mechanize’s public site remains live and still describes the company as building coding-agent environments and evaluations, which makes the transaction look more like a talent-and-technology arrangement than a standard acquisition.

Mechanize says it builds environments in which models perform real software-engineering work such as implementing features, deploying applications and debugging unfamiliar codebases. Graders score those attempts and the resulting signals can be used for reinforcement learning and evaluation. One public example, GBA Eval, asks an agent to build a Game Boy Advance emulator in Rust over a long task horizon. That expertise maps directly onto model midtraining and coding-agent improvement.

Earlier reporting put the talks at more than $1.5 billion for Mechanize technology and talent, though the final amount is still undisclosed. Mechanize had reportedly raised only $9.1 million earlier in 2026 at a $500 million valuation. Even without a confirmed closing price, the gap illustrates how aggressively frontier labs are pricing small teams that can create useful training environments, evaluation systems and feedback loops for coding models.

The structure resembles other AI talent-and-licensing deals where a large lab hires key staff and gains technology access without buying the whole legal entity. That can move scarce researchers and infrastructure into a model lab faster than a conventional acquisition, while leaving a smaller operating company behind. The exact Mechanize licensing terms have not been publicly confirmed, so the durable value Google receives beyond the staff moves remains uncertain.