OpenAI Five was a team of five neural-network "bots" that played the video game Dota 2, and it is the OpenAI system that earned the name "bot" before any chat product did. According to OpenAI's technical paper, it beat the reigning world champions, Team OG, two games to zero on April 13, 2019, after about 10 months of training, and it did so under a reduced version of the game.
Key takeaways
- OpenAI Five learned by self-play reinforcement learning, not from human game recordings, and trained from June 30, 2018 to April 22, 2019.
- The game was restricted: 17 of 117 heroes were allowed, and four items that split control across several units were removed.
- In a public online event on April 18 to 21, 2019, the system won 99.4 percent of 7,257 games, but about 3,140 of the wins came from opponents abandoning matches.
- The system had no language ability. It read structured game data and picked game actions, so it says little about chatbots.
What OpenAI Five was
The paper "Dota 2 with Large Scale Deep Reinforcement Learning," submitted to arXiv on December 13, 2019, lists 25 authors and describes the project in detail. Dota 2 is a five-against-five game with long matches, hidden information, and a very large number of possible actions: the authors say the model chose among 8,000 to 80,000 actions on an average step, depending on the hero. The agent acted on every fourth frame of a game that runs at 30 frames per second.
Each of the five heroes was controlled by a copy of the same policy, a recurrent neural network with about 159 million parameters, built mainly around a single-layer 4096-unit LSTM. The network did not look at the screen. It received the game engine's structured data, such as unit health and position, which the authors describe as the information a human player would see. The paper notes that this is an approximation, since some of what a person perceives is not in the data.
How it was trained
OpenAI Five learned through self-play: it played against copies of itself and was rewarded for winning and for intermediate game events. The paper says most games used the latest version of the policy and a minority used older ones, to keep the agent from forgetting how to beat earlier strategies. A single training run lasted from June 30, 2018 to April 22, 2019, and the paper reports that it used 770 petaflop/s-days of compute by the time of the OG match. The authors compare the scale to AlphaGo, citing batch sizes 50 to 150 times larger and a model about 20 times larger.
The project changed its game rules, observations, and network shape many times during those months. Instead of restarting each time, the team used a method they call surgery, which transforms an existing trained model to fit a new design. The paper counts over 20 surgeries. A separate run named Rerun, trained from scratch on the final setup between May 18, 2019 and June 12, 2019, took about two months and 150 petaflop/s-days. The authors estimate that retraining from scratch after each surgery would have stretched the project from 10 months to 40.
The game it played was not the full game
The paper states two limits plainly. The restrictions are summarized below with a third detail the authors also disclose.
| Aspect | What the paper says |
|---|---|
| Heroes | 17 of the 117 heroes in the normal pool were supported |
| Items | Illusion Rune, Helm of the Dominator, Manta Style, and Necronomicon were removed because they let one player control several units |
| Scripted logic | Item and ability purchase order, control of the courier unit, and which items heroes keep in reserve were handled by hand-written rules, not the policy |
| Input | Structured data from the game engine, not pixels |
These choices make the result narrower than a headline such as "AI beats humans at Dota 2" suggests. The authors say they believe the agent could improve without the scripted parts, and that it already reached superhuman performance before they removed them.
Reading the results with the caveats
The OG match was a best-of-three, and the paper reports that OpenAI Five won it 2-0. The authors add a warning against over-reading a single show match: one high-stakes series does not prove broad understanding, because machine learning systems often behave badly in unexpected situations. They ran OpenAI Five Arena from April 18 to 21, 2019 to test that. It played 3,193 teams in 7,257 games and won 99.4 percent, and 29 teams managed to beat it, for 42 lost games in total.
A footnote qualifies the headline figure. Human players often abandoned losing games, so the authors counted abandoned games as wins, and those made up 3,140 of the 7,215 wins. Reaction time is also reported: the agent reacted to game events in 217 milliseconds on average, against roughly 250 milliseconds for a typical human visual reaction in controlled settings, a figure the authors call a likely underestimate for real play.
What it means for the word "bot" today
OpenAI Five had no language component, so it does not belong in the same category as a chat assistant. Its relevance lies in method: it showed that reinforcement learning with very large batches and long training could handle a game with long horizons and hidden information. The authors list their own limits, including that surgery plateaued at a weaker skill than training from scratch and that stale data slowed training. Readers looking for a product called an OpenAI bot should check the company's current documentation, because names and offerings have changed since 2019.
FAQ
Did OpenAI Five learn from watching human players?
The paper describes training through self-play reinforcement learning. Human games were used to evaluate progress, with amateur and professional players and teams, not as the training source.
Was the match against OG played under normal Dota 2 rules?
No. The paper states that only 17 heroes were supported and four multi-unit items were removed, and some game mechanics were scripted.
Is OpenAI Five the same as ChatGPT?
No. OpenAI Five read structured game data and chose in-game actions. ChatGPT is a language-based assistant, and the paper on OpenAI Five does not involve text generation.
Is the 99.4 percent win rate a fair measure of skill?
It is a real result, but the authors note that abandoned games were counted as wins and that these were 3,140 of 7,215 wins. The OG series and the finding that 29 teams won at least one game give a fuller picture.
Related reading
- GPT-3 and the language models that followed
- Capabilities of OpenAI chat tools
- How AI text generators work
- The history of conversational interfaces
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Dota 2 with Large Scale Deep Reinforcement Learning, Berner et al., arXiv:1912.06680, OpenAI, 2019 (revised March 10, 2021 text)