Google may spend $1.5bn on a 35-person startup to catch up on AI coding

Google wants its AI to code better. It may have found a shortcut, and it runs through a startup of about 35 people.

The company is in talks with Mechanize, a San Francisco startup that trains AI agents to write software, Business Insider reported. The deal would be worth more than $1.5 billion, four people familiar with the talks said. It is still in progress, and the terms could change.

What Google would get is narrower than an acquisition. It would hire some of Mechanize’s staff and take a non-exclusive licence to its technology. The recruits would work on evaluating and developing Google’s models. Both companies declined to comment.

Why $1.5bn for 35 people

Mechanize is barely a year old. Three former researchers from Epoch AI, an institute that measures how capable models are, launched it in April 2025. Its chief executive, Tamay Besiroglu, co-founded Epoch. Earlier this year the startup said it had raised $9.1 million at a $500 million valuation, with backers including former GitHub chief Nat Friedman, Stripe’s Patrick Collison, and the podcaster Dwarkesh Patel.

Its stated mission is blunt: automate every job. It started with software because code can be graded.

That grading is the actual product. Coding agents learn inside training environments. Each one is a prompt, a working codebase, and a grader that decides whether the agent did the job. Those environments feed reinforcement learning and model evaluations. RuntimeWire reports that one engineer builds each task from idea to grading. That takes about a week. Most of the effort goes into making the grader fair, consistent, and hard for a model to cheat.

That is what Google is paying for. Not a consumer product, but the judgement to find where frontier coding models fail and build fair tests around it. A licence buys the finished work. Hiring the people keeps it coming as the models improve.

The playbook, again

Google has run this move before. Rather than buy a startup outright, it hires the key staff and licenses the technology. The structure sidesteps the antitrust scrutiny a full takeover attracts.

In July 2025 it took the Windsurf founders and a non-exclusive licence to their coding tool, a package reported at about $2.4 billion. Varun Mohan, Windsurf’s chief executive, now runs Google’s Antigravity coding platform. In 2024 Google rehired the Character AI co-founder Noam Shazeer and paid for rights to that startup’s technology. Shazeer has since left for OpenAI.

The Mechanize talks push the same playbook further down the supply chain. Windsurf gave Google a product developers use. Mechanize would give it the machinery to train the models underneath.

The timing is brutal

The deal surfaced in a bad week for Google’s AI. Days earlier, DeepMind boss Demis Hassabis stepped back and the chief scientist, Jeff Dean, left to start his own company. Google shares fell around 4% on that news. The lab had already been under visible strain over talent and delays.

The coding gap is the backdrop. Google’s flagship Gemini has slipped repeatedly on coding, while OpenAI and Anthropic pulled developers onto Codex and Claude Code. The fight has moved from the model to the tools around it, and rivals are building their own. Meta just launched its own coding agent. Buying Mechanize’s evaluation craft is a way to close the distance without waiting for the next Gemini.

What the headline number hides

The $1.5 billion figure needs care. RuntimeWire notes it is the value of the whole arrangement, not a purchase price or a fresh valuation. The money could be split across licensing fees, salaries, founders, and investors, leaving a smaller Mechanize still standing. Deedy Das called the talks an acquisition on X, but the underlying report describes a licence-and-hire deal that leaves the company independent.

There is a sharper reading doing the rounds. Google wrote the paper that started the large language model era, then lost the lead to OpenAI. Wccftech points to the awkward symmetry. Apple pays Google roughly $1 billion a year to use Gemini. Google may now pay $1.5 billion to a 35-person shop to make Gemini better at a job its rivals have already nailed.

For now it is talks, not a deal, and the number could move. But the direction is clear. The scarce thing in AI coding is no longer the model. It is the people who can build the tests it keeps failing.