Gemini 4 Minecraft win becomes a cost sample
The token estimate reframes the Minecraft win as a cost sample, not a service-level promise.
| Gemini 4 agents | |
|---|---|
| agents | four |
| run time | 27 hours |
| programs | 2,783 |
| deaths | 605 |
| difficulty | Easy |
| token cost | estimated $520 |
The program count comes from Tom's Hardware, which attributes 2,783 programs to four Gemini 4 agents in a 27-hour Minecraft run.
That count does not establish human setup, novelty, runnability, or independent verification. It leaves open how many programs were executed, rejected, or left in the final build, and whether the final build was stable.
The $520 token estimate is the real constraint. A 27-hour, 605-death run is a cost sample, not a price list or service-level promise. For a buyer, it shows agentic coding can spend hours and dollars to reach a win, but it leaves the model short of a reliable production worker.
What remains unmeasured is the token price per death, the Minecraft version, the agent harness, and whether another lab can reproduce the win at the same published token cost or a public replay.
The $520 estimate is the useful number, and the 2,783 programs remain unverified.
After deepmind.google. We did not report this. The pictures, if any, are theirs.
