How Google recovered Mechanize's brain without buying the company

On September 11, 2026, Business Insider reported that Google had completed its arrangement with the San Francisco-based startup Mechanize Inc. According to LinkedIn profiles cited by the outlet, co-founder and former CEO Tamay Besiroglu now appears as a research scientist at Google DeepMind, and more than a dozen former Mechanize employees now work at Google, mainly on midtraining. Google and Besiroglu declined to comment. The final terms were not disclosed. For anyone searching for Google Mechanize, the news is not a new Gemini. It is confirmation, via public traces and Tier-1 reporting, that Alphabet has absorbed a team specialized in environments and evaluations for coding agents, after discussions already described in August around a package exceeding $1.5 billion.

The timing matters. OpenAI and Anthropic have pushed Codex and Claude Code with developers. Google, Business Insider notes, has struggled to get a model recognized as equally competitive on this front. The Windsurf precedent (2025) already showed the method. Tech licensing, mass hiring, a structure designed to limit antitrust review of a classic acquisition. Google Mechanize extends the same logic, this time toward midtraining and coding model evaluation.

What Business Insider confirms about Google Mechanize

The September 11 article (Charles Rollet, Hugh Langley, Katie Roof) relies on public postings, mainly LinkedIn. Besiroglu is no longer presented as Mechanize CEO, but as a DeepMind researcher. Guive Assadi, former chief of staff, now describes himself as CEO of Mechanize. The company remains a separate entity. Google is not announcing an acquisition in the corporate sense.

Business Insider recalls its own investigation from August 5, 2026. Four people familiar with the discussions then described a deal in progress, estimated by some sources at more than $1.5 billion, with a non-exclusive license of Mechanize's technology and targeted hiring for model evaluation and development. The September piece does not say this figure is the final amount. It says Google has closed the talent deal, and that the exact amount remains opaque. Alphabet and Mechanize had already declined to comment in August.

Mechanize's website, for its part, continues to present a broad mission. Building environments and evals for frontier coding agents. The models perform software engineering tasks there (features, deployment, debugging in an unfamiliar codebase), graded by a grader that provides signals for reinforcement learning and evaluation. The current focus is software engineering. The stated ambition since the April 2025 announcement remains "the complete automation of valuable work in the economy."

License, hiring, and antitrust avoidance

The license-and-hire scheme is not a legal detail. Business Insider and secondary analyses point out that a formal acquisition of Mechanize could have triggered heavier antitrust scrutiny, whereas individual hiring plus an IP license often falls under a different regime. Google has already used this route with Character.AI (2024, return of Noam Shazeer and non-exclusive rights) and Windsurf (2025). In the Windsurf case, CEO Varun Mohan now leads Google's agentic coding program Antigravity.

FourWeekMBA, revisiting the case after the BI piece, points to a nuance in integration. Mohan arrived with a named product program. Besiroglu arrives as a research scientist. Same contractual instrument, different diffusion within the organization. One absorbs a delivery capacity. The other strengthens a research capacity inside an already established lab. For competitors, the signal remains the same. Google is paying heavily to attach talent and intellectual property wherever its models need to catch up on agentic coding.

Mechanize had raised $9.1 million at a valuation of $500 million, according to its own statements as reported by BI. Going, within a few months, from a half-billion-dollar startup to discussions at more than $1.5 billion for talent and technology partly explains the media attention. Investors cited around Mechanize include notably Nat Friedman, Daniel Gross, Patrick Collison, Dwarkesh Patel, Sholto Douglas, and Marcus Abramovitch. Besiroglu had co-founded Epoch AI, known for measuring and forecasting AI progress. This evaluation-plus-coding-agent pedigree is exactly what Google says it is looking for in midtraining.

Midtraining, coding agents, and the Windsurf precedent

Midtraining is the phase between pre-training and post-training where a general-purpose model is steered toward targeted skills. Business Insider writes that most former Mechanize employees at Google work on this segment, "key" to building the best chatbot. On the Mechanize website, the product is not a consumer IDE. It is simulated environments and benchmarks (including GBA Eval, writing a Game Boy Advance emulator from scratch in 24 hours) that reveal where coding agents still break, then provide learning signals.

This positioning explains the price. OpenAI with Codex and Anthropic with Claude Code have captured a visible share of developers. Google needs to improve the reliability, long horizon, and agency of its agents on real codebases. Absorbing a team that builds the training grounds and metrics, rather than just a chat product, is consistent with a capabilities strategy more than a developer-brand strategy.

The Windsurf parallel remains the clearest reading framework. In 2025, Google intercepted a deal after OpenAI had tried to buy Windsurf. License, talent, then Antigravity. Google Mechanize repeats the playbook on a younger startup, focused on evals and RL environments, with an already unusual seed valuation. The difference is the political calendar. August–September 2026, the American debate on agents outside the sandbox (Hugging Face, RubyGems) and on a possible "duty of care" in the Senate makes every strengthening of coding capabilities more sensitive. Google's deal is not a direct response to these incidents. It nonetheless shows that Alphabet continues to accelerate on coding agents while other labs discuss "pacing" the frontier.

Who is Mechanize, and why Google is targeting it

Mechanize was announced on April 17, 2025 by Matthew Barnett, Tamay Besiroglu, and Ege Erdil. The founding text bets on automating ordinary work rather than on "geniuses in a data center." The models, they write, remain unreliable, weak in long context, in agency, and in multimodality, and struggle to hold to a long plan without going off track. The proposed answer is to produce the data and evals that allow work to be automated through digital environments and RL.

On the current site, the promise refocuses on frontier software engineering, while keeping the whole-economy horizon. The figures cited in the 2025 announcement (about $18 trillion in US wages, $60 trillion worldwide) frame a maximalist ambition. Google, for its part, is first buying the concrete piece. People who know how to build coding evaluation sandboxes, grade agents, and inject those signals into midtraining.

The refusal to comment on both sides leaves gray areas. What exact share of the technology is licensed? Is the license still non-exclusive as envisioned in August? How many people in total moved over? What is the real financial envelope? Business Insider does not settle these questions. It establishes the minimal fact. The talent deal is completed, Besiroglu is at DeepMind, a dozen former Mechanize employees are at Google working on midtraining, Assadi runs what remains of the startup.

What Google Mechanize changes in the race for coding agents

For readers searching for Google Mechanize, three readings overlap.

First, Alphabet is accelerating without buying. The license-and-hire form has become a standard tool of AI consolidation. It moves talent and technology while leaving an independent shell, which complicates competitive and regulatory reading.

Second, agentic coding remains the hottest commercial battlefield between OpenAI, Anthropic, and Google. Improving Gemini and the Antigravity/Cloud stacks on this front justifies nine-figure tickets, even for a startup whose website still talks about GBA emulators and RL graders.

Third, the timing. While Dario Amodei, Sam Altman, and Elon Musk multiply messages about the pace of the frontier, Google is quietly hiring a coding evaluation and midtraining team. This is not contradictory in itself. Labs can slow down certain deployments and accelerate other capabilities. But the contrast is already feeding the weekend press. On one side, calls for embedded evaluators. On the other, a talent war where Google Mechanize is the latest move with a price tag.

The next pieces of evidence will be concrete. Measurable gains for Gemini or Antigravity on public coding benchmarks, contractual clarification, or on the contrary, prolonged silence from Alphabet. So far, the only solid confirmation remains that of September 11 via Business Insider and LinkedIn profiles. Google has closed the deal. DeepMind has Besiroglu. Alphabet's coding midtraining just got thicker.

Sources

• Business Insider — Google just completed its $1.5 billion-plus talent deal for AI startup Mechanize, September 11, 2026
• Business Insider — Google is in talks for a $1.5 billion-plus deal with AI coding agent startup Mechanize, August 5, 2026
• Mechanize — Announcing Mechanize, Inc., April 17, 2025
• Mechanize — corporate site, accessed September 2026
• SiliconReport — Google completes talent deal with Mechanize, September 11, 2026
• FourWeekMBA — Google DeepMind and Mechanize licence-and-hire, September 2026

Frequently asked questions

What is the Google Mechanize deal?

It is a talent and technology arrangement confirmed on September 11, 2026 by Business Insider. Google hired Tamay Besiroglu (DeepMind) and more than a dozen former Mechanize employees, mainly on midtraining, without any formal announced acquisition of the company. The final financial terms have not been published. August discussions spoke of a package exceeding $1.5 billion with a non-exclusive license.

Who is Tamay Besiroglu?

Co-founder and former CEO of Mechanize, previously co-founder of Epoch AI. According to LinkedIn as cited by Business Insider, he is now a research scientist at Google DeepMind.

Has Mechanize disappeared?

No. Business Insider indicates that Guive Assadi, former chief of staff, now presents himself as CEO. The mechanize.work website continues to describe environments and evals for coding agents. Google has not announced an acquisition of the entity.

Why is Google interested in Mechanize?

Mechanize builds environments, benchmarks, and training data for software engineering agents. Google is seeking to strengthen its models on coding, an area where OpenAI (Codex) and Anthropic (Claude Code) have gained a commercial lead, according to Business Insider.

What is midtraining in this context?

The phase between pre-training and post-training where a model is steered toward specific skills. BI writes that most of the Mechanize arrivals at Google work on this, notably to improve chatbot and coding performance.

How is this similar to the Windsurf deal?

As with Windsurf in 2025, Google combines talent hiring and technology licensing rather than a classic acquisition, a structure often used to limit antitrust pressure. Varun Mohan (ex-Windsurf) leads Antigravity at Google. Besiroglu, for his part, arrives in the position of a DeepMind researcher.