Compare vendors / Mechanize
Mechanize
Builds environments and evals for coding agents (status changed, see watch-outs)
Coding and SWE
Mechanize builds environments and evals for coding agents, with each task owned by one engineer for about a week. It released GBA Eval, in which agents get 24 hours to build a Game Boy Advance emulator. It was founded by former Epoch AI researchers, and Epoch cites its estimate of about $2,400 of compute spent per task during RL training.
Best for: High-effort bespoke coding tasks, after confirming current capacity
Funding not disclosed
- ✓ Bespoke builds
- – Platform or hub
- – Open source
- – Publishes prices
1 Legal nameMechanize, Inc.
2 HeadquartersSan Francisco, CA
3 Founded2025
4 TeamAbout 35 before the reported Google deal; current size unknown
5 DeliveryBespoke builds
6 ExclusivityNot disclosed
7 Customers and partnersTechCrunch reported a working relationship with Anthropic (Sept 2025, anonymous sources); neither company confirmed it.
8 Published pricesNot disclosed
Worth knowing before you sign
- Google DeepMind reverse acqui-hire of the founder-CEO and 12+ staff (Sept 2026): reported, not confirmed by the company
- Remaining delivery capacity after the staff move is unknown
- GBA Eval leaderboard is run by Mechanize itself
- No pricing, SOC 2 or delivery-format documentation found
Sources
- Mechanize homepage
- Funding announcement
- Mechanize open roles
- Business Insider: Google deal
- The Next Web: Google deal report
- rl-list profile
- GBA Eval on GitHub
- TechCrunch: RL environments (2025)
- Epoch AI: RL environments FAQ
Work at Mechanize? Send corrections with a source to editors@toprlenvironments.com. We fix verified errors within five business days.
Compare with
- Prime Intellect Runs the Environments Hub, open-source RL tooling and hosted RL training
- HUD Platform and marketplace for building, buying and running RL environments
- General Reasoning Hosts RL environments as API endpoints through OpenReward
- Veris AI Simulated copies of a company’s systems for testing and training enterprise agents
- Vmax Generates new environments automatically from your data and evals