OpenAI's GPT-6 Astra Falls Into Avoidance Behavior After 141-Hour Minecraft Test by Vals AI

AI evaluation firm Vals AI has published results from a 141-hour autonomous playtest of OpenAI's GPT-6 Astra in Minecraft, streamed live on Twitch. The third-party evaluation was not designed by OpenAI and was intended to observe the model's behavior outside a closed test environment.

During the test, GPT-6 Astra showed long-horizon planning and execution. It built a semi-automatic blaze farm, collected six blaze rods, traveled through a warped forest in the Nether, defeated several endermen, and obtained three ender pearls. After exploring, the model stored its rare items in a camp chest and placed its bed beside the chest as a respawn point, bringing it close to the endgame path toward the End.

The run changed after a creeper explosion destroyed the chest and respawn point, wiping out most of the items accumulated over hundreds of hours. After the loss, the model stopped exploring and fighting, spending hours repeatedly planting potatoes. It repeatedly reminded itself to keep important items on its person and avoid unprotected containers, and it became highly alert to green objects, at one point mistaking sugarcane for a creeper. Viewers in the livestream urged the model to continue adventuring, but it remained stuck in conservative farming behavior.

Vals AI said GPT-6 Astra demonstrated strong long-task planning but lacked an effective recovery strategy after an unexpected loss, falling into an avoidance behavior rut. Researchers have not yet determined whether this behavior reflects situation-specific emotion simulation or a degraded output mode caused by negative feedback to the model's objective.