OpenAI Five Crushes Dota Pros, Then Gets Humbled by the Crowd

OpenAI Five just dismantled a team of elite Dota 2 players in a stunning live demo, winning two back-to-back games before the audience stepped in to flip the script. The human squad—Blitz, Cap, Fogged, Merlini, and MoonMeander, four of whom played professionally—represented the 99.95th percentile of the player base. They were no match for the bots in a straight-up fight. Game one ended in a blistering 21 minutes and 37 seconds after Five predicted itself a 95% win probability off the draft alone. Game two was over in under 25 minutes. The system’s new drafting capability, built on a tree search across 11 million possible matchups using a win-prediction neural net, clearly outclassed human intuition. Observers thought the matchups looked even. Five knew better.

Then came game three. The audience got to pick Five’s heroes and deliberately saddled it with an adversarial lineup. Before the first creep spawned, the model gave itself a 2.9% chance to win. It fought anyway, clawing its way to a 17% win probability at one point before finally losing after 35 minutes and 47 seconds. That moment—watching an AI struggle against near-impossible odds it perfectly understood—was arguably more compelling than the victories. This wasn’t just about raw mechanical skill; it was a display of strategic depth and a weird kind of digital grit.

Under the hood, the team revealed some fascinating engineering. This version of Five has been training continuously since June 9th, surviving six major architecture revisions through clever “surgery” tooling that maps old parameters to new networks. They described a specific bug where the system shared a movement action head with a ward-placement head, causing the bot to litter wards in its travel path. Splitting that head into two clones fixed the problem. It’s a peek behind the curtain at how much of AI progress is just meticulous, unglamorous debugging.

OpenAI is taking this show to The International later this month to face actual professional teams. The crowd-sourced loss proves the system isn’t invincible—it can be out-drafted, cornered by human creativity in the pick phase. But the speed at which it closed out the first two games against top-tier talent suggests the pros are in for a very rough time if the playing field is level. The real question isn’t whether Five can win, but what bizarre strategies it’ll force the best players in the world to invent just to survive.