GPT-6 Astra’s game tests highlight response-latency limits

GPT-6 Astra’s Minecraft behavior was described as farming potatoes for several hours after a creeper destroyed its chest, with the “depression” framing remaining an interpretation rather than an established internal state. In single-player Minecraft, pausing may make the test somewhat easier.

The model reportedly handled games that allow long waits between actions, including Slay the Spire and Balatro; its Balatro result was described as impressive. Games requiring high throughput and high actions per minute are considered a more difficult test because response speed matters.

A possible future benchmark would be for Astra to win Dota 2 without training on Dota 2. A competitive lap time at Laguna Seca was also suggested as another potential test, not a reported result.