Better AI, with games
Games are compact worlds of skill and physics. We turn them into training materials that make models more capable.
Games test intelligence
- Winning takes planning, memory, spatial reasoning, and negotiation
- Every move comes with a verifiable signal of what good looks like
- Models that learn those skills in play carry them into real world, multimodal work
Interactive learning environments
High volume from watching real people play
Published research
Peer-reviewed papers and technical write-ups from our work.
Keep Going
GPT-6 Astra in Codex and Claude Opus 5.5 in Claude Code each write a StarCraft bot and climb a ladder of human-written bots, live for 48 hours.
Verification Is the Bottleneck
We had early access to TypeSafe’s Jev and found it incredibly powerful as a verification tool: $160 per million graded answers against $33,000 for Claude Fable 5.1.
Why do we do what we do
AI improves fastest where the goal can be checked. Games make far more goals checkable, and their rules decide what a model becomes.
Let's talk
Whether you are training frontier models or publishing games, we would love to hear from you.