OpenAI Five crushed Dota 2 champs using 2M frames per second

On April 13th, 2019, OpenAI Five didn’t just win a game — it dismantled Team OG, the reigning Dota 2 world champions, in a live match. That’s not an exhibition stunt. It’s the moment a reinforcement learning system crossed into superhuman territory on a problem that actually looks like the real world: long time horizons, fog of war, and continuous control. Dota 2 isn’t chess. You can’t brute-force it. You have to guess what opponents are hiding, coordinate with teammates, and make decisions that won’t pay off for another ten minutes. That’s an entirely different class of difficulty.

The scale is what makes your head spin. The system ingested roughly 2 million frames every 2 seconds during training — a throughput that’s frankly absurd. To put that in perspective, that’s about 180 years of human gameplay experience consumed every single day. The engineering team built a distributed training infrastructure and a set of continuous training tools that let them iterate on the system across 10 grueling months. This wasn’t some academic proof of concept that ran in a weekend on a GPU cluster. It was an industrial-grade grind.

Here’s the part that should make you sit up straight: they didn’t hardcode strategies or teach the AI how to play. OpenAI Five learned through self-play reinforcement learning, starting from zero and playing against copies of itself billions of times. No human demonstrations. No curated playbook. The same general approach that learned to walk in simulation and play Atari games now handles a team-based strategy game where individual matches stretch past 45 minutes and the state space is effectively infinite. That generality is the whole ballgame. If the method works here, it works on problems where we can’t even write down the rules clearly.

Skepticism is healthy, and yes, the match had some constraints — a limited hero pool, no illusions, no summons. OG didn’t get to play their full game. But dismissing the result because it wasn’t a completely unrestricted match misses the point. The system demonstrated something fundamental: self-play RL can produce long-horizon strategic reasoning that beats the best humans alive. The question isn’t whether it plays Dota 2 perfectly. The question is what else looks like Dota 2. Supply chains. Energy grids. Military logistics. That’s where this gets interesting.