From Mexico — I’m catching up on the Google–Mechanize talent close that Business Insider and Superpower Daily both clocked around Sep 11. Public LinkedIn moves show the San Francisco startup’s people are already inside Google. Co-founder and ex-CEO Tamay Besiroglu is listed as a research scientist at Google DeepMind. Over a dozen other former Mechanize folks are at Google too — mostly on midtraining, the messy middle that shapes how chatbots actually learn to code.
One number keeps getting misread, so I’m saying it slow: Google had been in talks for a $1.5 billion-plus deal for Mechanize tech and talent, BI previously reported. Final terms were not disclosed. That $1.5B+ figure is the talks headline, not a confirmed purchase price. Google and Besiroglu declined to comment. Former chief of staff Guive Assadi now identifies as Mechanize CEO on LinkedIn — the company shell is still there.
Midtraining people, coding evals
Mechanize builds environments and graders for frontier coding agents — the kind of RL + eval stack you use when short prompt benchmarks stop telling the truth. Their public showcase is GBA Eval (May 2026): task a model to write a Game Boy Advance emulator from scratch in Rust, compile it to WebAssembly, then grade with deterministic gameplay replay against a lightly modified Mesen2 reference. That’s sustained engineering, not a leetcode spritz. Superpower Daily ties that specialty straight to why midtraining hired them.
Earlier this year Mechanize said it raised $9.1 million at a $500 million valuation, per BI and Superpower Daily. Besiroglu previously cofounded Epoch AI (benchmarks / testing). BI also frames the structure as a talent deal vs a full acquisition — same pattern as Google’s Windsurf talent move, with Varun Mohan later leading agentic coding work branded Antigravity. I’m not inventing a close price. I’m reading the LinkedIn trail and the eval product.
What I’m watching from Mexico
Launch context, not the hook: Mechanize opened in April 2025 with a full-automation mission and white-collar focus, backed by Nat Friedman / Daniel Gross and others, as TechCrunch covered then. Today’s story is narrower — Google just absorbed the eval/midtraining muscle. For the primary GBA write-up, read Introducing GBA Eval on mechanize.work.
From Mexico, the bookmark for me is the midtraining assignment plus that grading philosophy — not the inflated billion-dollar rumor. If Google’s coding models get sharper over the next quarters, this talent path is one place I’ll look. Until then: deal completed, terms private, Assadi still on the letterhead, and the public evidence is people + profiles, not a press-release price tag. Sources: Business Insider, Superpower Daily, Mechanize’s own site.
