Google Pursues $1.5 Billion Deal with 35-Person AI Coding Startup Mechanize to Boost Model Performance

Google Targets Mechanize in $1.5 Billion Licensing and Talent Deal

Google is in active discussions with Mechanize, a San Francisco-based AI coding startup, over a deal valued at more than $1.5 billion1

1

2

2

. The proposed Google Mechanize deal would grant the tech giant a non-exclusive license to the startup's technology while recruiting several of its key employees to work on evaluating and developing Google's AI models' coding abilities1

1

. Four people familiar with the conversations confirmed the talks are ongoing, though terms could still change2

2

. Neither Google nor Mechanize has commented on the negotiations3

3

.

What makes this deal remarkable is Mechanize's size and age. The AI coding startup employs roughly 35 people and launched just over a year ago in April 20251

1

3

3

. Three former researchers from Epoch AI, including CEO Tamay Besiroglu who co-founded that institute, established Mechanize with a blunt mission: automate every job1

1

. Earlier this year, the startup raised $9.1 million at a $500 million post-money valuation from notable backers including former GitHub CEO Nat Friedman, Stripe CEO Patrick Collison, and podcaster Dwarkesh Patel1

1

2

2

.

Why Google Needs Mechanize's Evaluation Technology

Mechanize's core product centers on training AI agents to write software through specialized grading systems and simulated environments2

2

. The startup builds environments and evaluations for AI coding agents, enabling AI systems for coding to carry out software engineering tasks within these environments while scoring performance to support reinforcement learning and model evaluations3

3

. Each training environment consists of a prompt, a working codebase, and a grader that determines whether the agent completed the task successfully1

1

.

What Google is actually purchasing is the expertise to identify where frontier coding models fail and construct fair, consistent tests around those weaknesses1

1

. One engineer typically spends about a week building each task from concept to grading, with most effort dedicated to making the grader fair, consistent, and resistant to model manipulation1

1

. Technology licensing gives Google access to finished work, while talent acquisition ensures continued development as models improve1

1

.

Google's Growing Gap in AI Coding Performance

Google faces mounting pressure in AI coding as its flagship Gemini models have slipped repeatedly in coding performance while OpenAI and Anthropic attract developers with Codex and Claude Code1

1

. The competition has shifted from raw model capabilities to the tools and infrastructure surrounding them, with rivals building their own complete ecosystems1

1

. Meta recently launched its own coding agent, further intensifying the race1

1

.

Google's coding agents are generally considered inferior to those of Anthropic and OpenAI, which have seen strong enterprise adoption2

2

. Acquiring Mechanize's evaluation capabilities represents a path to close this distance without waiting for the next Gemini release1

1

. The deal surfaces during a particularly difficult period for Google's AI division, coming days after DeepMind boss Demis Hassabis stepped back and chief scientist Jeff Dean left to start his own company1

1

. Google shares fell approximately 4% following that leadership upheaval1

1

.

Google's Established Corporate Strategy for AI Talent

This approach reflects Google's established corporate strategy of structuring hybrid transactions rather than outright acquisitions to avoid antitrust scrutiny1

1

2

2

. In July 2025, Google recruited top executives and researchers from Windsurf and secured a non-exclusive license to their coding tool in a package reportedly worth $2.4 billion1

1

3

3

. Former Windsurf CEO Varun Mohan now leads Google's Antigravity coding platform1

1

2

2

.

In 2024, Google rehired Character AI co-founder Noam Shazeer and paid for rights to that startup's technology, though Shazeer has since departed for OpenAI1

1

2

2

. The Mechanize discussions push this playbook deeper into the supply chain—while Windsurf provided a product developers use directly, Mechanize would supply the underlying machinery to train the models themselves1

1

.

Understanding the $1.5 Billion Valuation Structure

The $1.5 billion figure represents the total value of the arrangement rather than a straightforward purchase price or company valuation1

1

. This sum could be distributed across licensing fees, salaries, payments to founders, and investor returns, potentially leaving a smaller but independent Mechanize still operating1

1

. While some observers have characterized the talks as an acquisition, the underlying reports describe a license-and-hire structure that preserves the company's independence1

1

.

The discussions highlight two critical trends shaping the AI industry: coding has emerged as one of the most lucrative applications of AI models, and major technology firms are willing to pay substantial sums to secure promising AI developer talent2

2

. AI coding assistants enable smaller teams to automate tasks performed by knowledge workers that would otherwise require additional engineers and significant costs, while cutting development time for businesses to bring products to market faster3

3

.

Mechanize's ambitions extend well beyond coding. The startup aims to create simulated environments and evaluations capturing the full scope of what people do at their jobs, including using computers, completing long-horizon tasks lacking clear success criteria, coordinating with others, and reprioritizing when facing obstacles and interruptions3

3

. As the company states on its website, current focus remains on software engineering, but the long-term goal is full automation of valuable work across the economy2

2

. For now, the talks remain ongoing and the final terms could shift, but the direction is clear: the scarcest resource in AI coding is no longer the model itself, but the people who can build the tests that expose where those models still fail1

1

.