It took 8 minutes and 43 seconds for an AI to defeat the Ender Dragon in Minecraft. On September 20, engineer Ronak Malde published the experimental results of a Minecraft agent combining OpenAI's "GPT-6 Astra" and TypeSafe AI's "Jev."

Jev + Astra beats the Ender Dragon in Minecraft in 8 minutes 43 seconds! ⏱️ Cost less than $1 ($0.01 Jev, $0.96 Astra) I open sourced the code and explain the harness setup below. This type of movement is only possible with Jev's near instant decisionmaking, and some… pic.twitter.com/daKubJ8zpU — Ronak Malde (@rronak_) September 20, 2026

Jev is not an AI for writing text like ChatGPT, but rather an AI designed to quickly determine "which action should be chosen now." For example, it answers questions like "Which of A, B, or C to choose?", "Is it dangerous or safe?", or "What is the importance score?" not with text, but with choices or scores. It's, so to speak, a "fast if-statement that understands the meaning of words."

Diogo Almeida, founder and CEO of TypeSafe AI, which developed Jev, is a former OpenAI researcher who was involved in research such as InstructGPT and RLHF, which later led to ChatGPT. After leaving OpenAI, Almeida tackled the question of "why AI has become so smart in conversation, yet job automation hasn't progressed as much," and released Jev on September 15.

The Minecraft experiment clearly demonstrates Jev's characteristics. Starting a new world in Minecraft Java Edition 1.16.5 with an empty inventory, it traveled through the Nether to the End. Finally, it defeated the Ender Dragon using bed explosions. During execution, Jev was called 131 times, and Astra 35 times. The model usage fees were approximately $0.01 for Jev and $0.96 for Astra.

What's interesting is that the two AIs are assigned completely different roles.

Astra is responsible for high-level planning, such as "what should be achieved now" and "where to go next." Jev, on the other hand, receives the current location, inventory, surrounding conditions, and the goals set by Astra, then selects the action to perform at that moment from candidates like "move," "dig," "pick up," "craft," or "attack."

In other words, Astra thinks about "how to strategize," while Jev decides "what to do right now."

In light of the concept that divides human thought into "intuitive System 1" and "deliberative System 2," Astra plays a role similar to System 2, which considers the strategy, while Jev plays a role similar to System 1, which decides "what to do now."

In terms of implementation, the game does not pause while Astra updates its plans. Astra thinks about the next strategy in the background, and during that time, Jev continues to select actions according to the current goal.

Actual movement, mining, and other operations are handled by "Mineflayer," a bot library for Minecraft. This creates a division of labor: "Astra, who thinks long-term," "Jev, who makes instant decisions," and "conventional software that performs the actual operations."

To illustrate with a human player analogy, Astra would be the one thinking, "To defeat the Ender Dragon, first I'll gear up and head to the End." Jev would be the one deciding, "I'll open the chest in front of me." And Mineflayer would be the one actually walking to the chest.

However, this experiment does not mean that the AI conquered Minecraft on its first try under the same conditions as a human. The difficulty was Peaceful, the seed and route were pre-selected, and known coordinates were utilized. Furthermore, the AI is not directly operating the keyboard by looking at the screen, but rather using structured data obtained from within Minecraft.

The significance of this experiment lies in the fact that it demonstrated that a new, different kind of value can emerge from the combination of a "thinking LLM" and an "instantly deciding Jev."

Until now, AI agents often relied on a single large LLM for everything from long-term planning to minute decisions like "which button to press next." However, advanced text generation capabilities are not necessary for every decision.

Jev acts as a "decision-maker" that connects the plans conceived by an LLM to concrete operations in real-world software and tools. The successful defeat of the Ender Dragon in Minecraft serves as a clear example illustrating the purpose of this new type of AI, Jev.