Breaking

OpenAI's GPT-6 Astra Reaches Minecraft Record Before Stalling After a Creeper Loss

September 16, 2026 at 12:13 EDT

Foundation Models
AI Agents

OpenAI's GPT-6 Astra reached a new high-water mark in a 141-hour Minecraft computer-use run, then stalled after a Creeper destroyed the chest and bed holding key progress items. The episode turns an unusually public game test into a useful snapshot of both long-horizon agent gains and still-fragile recovery behavior in frontier models.

September 2026 · GPT-6 Astra

A Minecraft milestone, then a recovery failure

A 141-hour run showed frontier-model progress, then a single Creeper loss pushed the agent into hours of low-value recovery behavior.

141h

Minecraft run

Blaze Rods collected

Ender Pearls secured

72.6%

OSWorld 2.0 score

Computer-use speed comparison

Reported OSWorld 2.0 time per task. Each block represents about five minutes; lower is faster.

How the run turned

Progress

Blaze farm, rods, forest and pearls

Creeper loss

Chest and bed destroyed

Recovery loop

Hours of potato farming and threat checks

Positive read

The agent sustained a complex run long enough to reach late-game objectives that previous AI systems had not reached in the test.

Critical read

The same run exposed brittle recovery after state loss, with the agent shifting into low-value repetition instead of quickly rebuilding.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.