OpenAI Five Finals

OpenAI just posted a short note confirming the finale: April 13, 11:30am Pacific, and that's it for live OpenAI Five events. No sequel season, no rematch tour. Just one more show and the curtain drops on a project that's been quietly reshaping how people think about reinforcement learning at scale.

OpenAI Five isn't a single algorithm tweak or a clever trick. It's a team of five neural networks trained to play Dota 2, a game with more moving pieces than chess and Go combined — hidden information, real-time decisions, five-on-five coordination, matches that stretch for 45 minutes or more. Getting an AI to hold its own in that environment meant throwing an enormous amount of compute at self-play, letting the bots battle themselves millions of times and slowly get less terrible.

What made this project matter beyond the scoreboard was the visibility. OpenAI didn't train Five in a lab and publish a paper six months later. They put it on stage, against actual professional players, in front of a live audience, repeatedly. That's a much riskier way to show off research — losses are public, glitches are public, everything is public. And they did it anyway, treating each match less like a PR stunt and more like a stress test for whether their training methods actually generalized to something messy and human-designed.

Ending the live events doesn't mean the research stops mattering. It means OpenAI is closing this particular chapter of showing the work in real time. Whatever comes next — and something always comes next with this lab — will likely lean on lessons learned from watching Five stumble, adapt, and eventually beat opponents who'd spent years mastering the game professionally.