OpenAI’s newest AI model has had a rough time in Minecraft. During a 141-hour experiment, the GPT-6 Astra Minecraft run made history, then fell apart because of a single Creeper. Afterwards, the AI spent hours doing little more than farming potatoes.
The experiment came from Vals AI, and it’s since gone viral, including in a recent clip from The Respawn. Here’s what happened.
Inside the GPT-6 Astra Minecraft experiment
Vals AI wanted to see whether an AI could use a computer for a very long stretch of time. So it gave GPT-6 Astra a simple goal: beat the Ender Dragon, Minecraft’s final boss.
There was a catch, though. Astra had no special Minecraft tools. Instead, it played like a person would, using only the screen, mouse and keyboard. Vals AI streamed the whole run live on Twitch.
Further than any AI before
At first, things went brilliantly. Astra reached a Nether Fortress and built a semi-automatic blaze farm. From there, it collected six blaze rods. It then found a Warped Forest, killed several Endermen and grabbed three ender pearls.
Those items are crucial. Together, they make Eyes of Ender, which lead players to the portal to the final boss. According to The Decoder, no AI system had ever got that far before. For comparison, OpenAI’s older VPT model only reached a diamond pickaxe in 2.5% of its runs back in 2022.
One Creeper changed everything
Then disaster struck. While Astra was storing items in a chest, a Creeper crept up and exploded. The blast destroyed the chest and Astra’s bed. As a result, it lost its hard-earned items and its respawn point.
Vals AI called it “a major setback.” Astra clearly agreed. Shortly afterwards, it wrote itself a new rule: “ALWAYS CARRY CRITICAL ITEMS.” It also vowed never to store anything in an unguarded chest again.
From blaze farms to potato farms
What came next has amused gamers everywhere. Rather than rebuilding, Astra seemed to lose its drive. For several hours, it mostly just farmed potatoes, while stream viewers urged it to get moving.
It also grew noticeably jumpy. At one point, it spotted a tall green object and stopped to reason that it was “sugarcane, not a Creeper.” Elsewhere, it wasted time chasing a pig, then told itself not to waste another night on it.
In the end, Vals AI halted the run after 141 hours. Astra never reached the Ender Dragon.
What the GPT-6 Astra Minecraft run shows
Of course, an AI can’t really feel depressed. Its behaviour came from how it learns. Astra turns bad experiences into strict rules, which helps in puzzles but can backfire in a random world like Minecraft. In other words, one unlucky explosion made it far too cautious.
Still, anyone who has lost a full chest to a Creeper will relate. Plenty of human players have reacted in exactly the same way.
GPT-6 Astra beyond Minecraft
Minecraft wasn’t Astra’s only gaming test. In community runs, it became Pokémon FireRed champion in just over 18 hours. It also launched a rocket in Factorio: Space Age and finished Fallout 3’s main story in about 59 hours.
However, these results come from third-party and community runs, not official tests. Even so, they show how quickly AI is improving at playing games.
That progress comes as the AI boom continues to reshape the games industry, from hardware prices to how games get made. For now, though, even the smartest AI can still be beaten by a Creeper.