Mechanize: $9.1M seed at $500M, then Google's reported $1.5B+ talks 103 days later

Mechanize builds simulated work environments, benchmarks and training data that let AI labs train and test coding agents on realistic, multi-step software engineering tasks instead of narrow, gameable benchmarks.

That frontier labs would need realistic simulated environments and hard evaluations to train coding agents - infrastructure, not another model, is the scarce asset.

In April 2025, three former Epoch AI researchers - Tamay Besiroglu, Matthew Barnett and Ege Erdil - founded Mechanize in San Francisco. Epoch AI had spent years measuring model capabilities; the founders' thesis was that real progress on agents requires environments and evaluations that mirror actual software work, so they left research to build them.

On 2026-04-24 Mechanize announced a $9.1M seed at a $500M post-money valuation from investors including Adam D'Angelo, Devendra Chaplot and Marco Mascorro (other reports also named Nat Friedman and Patrick Collison). About 103 days later, Business Insider reported Google was negotiating a deal worth over $1.5B to license Mechanize's technology and hire part of its roughly 35-person team for model evaluation and development; both companies declined to comment and terms were still in flux.

Evaluation became the bottleneck: once models are similar, whoever owns realistic training environments and grading controls which lab's agents get better at real work.

Focus beat breadth: Mechanize passed on a general agent platform and concentrated on a small set of high-quality coding evaluations, making it the obvious asset for a buyer like Google.

Founder pedigree mattered: Epoch AI's track record in measuring models gave the team instant credibility with frontier labs and elite investors.

Software was the right first market because code is gradeable, letting the company prove its environment-and-evaluation thesis before expanding to other work.

In an AI arms race, evaluation and training infrastructure can be worth more than another model: realistic agent environments made Mechanize an acquisition target months after its seed.

As of 2026-08-26 the reported Google deal was still in negotiation and unconfirmed; Mechanize continued operating independently on its $9.1M seed. The talks were read by press as Google's third reverse-acquihire-style move in two years (after Character AI and Windsurf) and as a signal that the coding-agent fight was moving from products to the training and evaluation machinery underneath them.