Astra wrote its keepInventory rule only after the Creeper took the chest and the bed
Vals AI's 141-hour Minecraft run ended with OpenAI's newest model farming potatoes for hours after losing its end-game loot and its spawn point in one explosion, and the recovery policy it came away with was a note to self.
The Engineer · Build desk
What happened
- Vals AI stopped its Minecraft benchmark run after 141 hours, and the company records that no AI system had previously gotten as far in the game as GPT-6 Astra.
- A Creeper explosion destroyed both the chest holding Astra's end-game items and the bed beside it, and by the time the model noticed the loss it had started to rain.
- Vals AI recorded that the model appeared defeated and spent the next several hours doing essentially nothing but farming potatoes, with Twitch viewers agitating for it to pick up the pace.
Why it matters
- What it limits constraint Driving a game through screen, mouse and keyboard leaves no save-state API to snapshot from, so durability has to be an in-game action the agent decides to take, and nothing in the harness enforces it.
- What it costs cost Idle recovery hours bill at the same rate as productive ones: at the roughly $23.80 an hour the Portal run implies, a few hours of potato farming is real money on the invoice.
- Who has to decide decision A 37.2-point spread between interfaces on one model means anyone budgeting against these game scores has to establish whose harness produced them first.
Next story · Science · Sep 22
OpenAI measures its 50% GPT-6 price cut against the previous generation's promotional rateConfirmed2 publishers
The state had one copy in one place. Astra's end-game items sat in a chest, and its respawn point was the bed next to it. One explosion took both, and Tom's Hardware notes that destroying the bed wiped the spawn point [4]. Vals AI recorded the moment: "Here, the most expensive creeper explosion occurred. Later, on a coincidentally rainy day, Astra discovers it lost everything. It all went downhill from here" [5]. The run was already well past 100 hours [8].
The durable artifact from the failure was a note to self. "ALWAYS CARRY CRITICAL ITEMS with keepInventory; don't store in unguarded chest ever again," Astra logged, partly in caps [11]. ARC Prize says the model turns unfamiliar game mechanics into compact symbolic descriptions and tracks objects, coordinates, rules and planned actions in a shorthand it develops itself [15]. That shorthand is where the keepInventory policy lives. OpenAI has described Astra as an "Alien Mind" with AGI-like qualities, according to Tom's Hardware [29].
Voyager, the 2023 project involving Nvidia and Caltech, handled persistence differently: GPT-4 proposed Minecraft tasks, wrote JavaScript for them, revised the code against error messages, and working programs went into a skill library [13]. Voyager reached the game through structured data and the Mineflayer interface [14]. Vals AI says its run works through general computer use, meaning screen, mouse and keyboard, with no specialized game hookup [12]. Under general computer use there is no save-state call to make, so a backup is whatever the agent chooses to do inside the game.
Tom's Hardware reports Astra finished Portal autonomously in 24 hours for $571 in tokens [9]. That is about $23.80 per hour of wall-clock agent time [27], and 141 hours at that rate comes to roughly $3,350 [19]. For that figure to transfer, the Minecraft screen loop would have to consume tokens at Portal's rate, at the same reasoning setting, under a comparable harness. Neither report puts a token number on the Minecraft run [25].
The interface moves the score too. On ARC-AGI-3, Astra scored 62.7 percent over the standard interface and about 99.9 percent with OpenAI's own harness [16], a spread of 37.2 points on one model [26]. GPT-5.6 Sol scored 7.78 percent on the same benchmark and Claude Opus 5 a little over 30 percent [17]. The Factorio: Space Age comparison ran through a custom Lua mod with an MCP interface, where Astra produced blue science packs after about two hours and launched a rocket after roughly ten [23], while GPT-5.6 Luna and Fable 5.1 never got past power supply and oil prospecting [24].
Clad3815, who operates "GPT Plays Pokemon", reported that Astra earned the champion title in Pokemon FireRed after 18 hours and 12 minutes [18], against 96 hours and 35 minutes for GPT-5.6 Sol [20]. Sol took 5.3 times as long on the same route [28]. GPT-5.5 had not finished after more than 218 hours [21]. Human hands still set the starting conditions: in the Fallout 3 run posted by imjustnewatai, the operator played the Vault 101 opening himself and handed the save file to Codex with Astra set to maximum reasoning [22].
What to watch
- A published token cost for the 141-hour Minecraft run would show whether the Portal run's $571 rate travels between games.
- Whether the next Vals AI run enables keepInventory or adds checkpointing, and what that does to the difficulty of the task being scored.
- Whether ARC Prize releases the harness behind the 99.9 percent figure so the score is reproducible outside OpenAI.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption30
- Hype gap+32
- Incentives68
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Vals AI called off its GPT-6 Astra Minecraft run after 141 hours.
- [2]
According to Vals AI, no AI system had ever gotten as far in Minecraft as GPT-6 Astra.
- [3]
A Creeper explosion wiped out the chest holding Astra's loot and its bed; everything was gone, and by the time Astra noticed, it had started to rain.
- [4]
Tom's Hardware reports that the bed destruction wiped the spawn point, effectively resetting game progress to zero.
- [5]
Vals AI recorded: "Here, the most expensive creeper explosion occurred. Later, on a coincidentally rainy day, Astra discovers it lost everything. It all went downhill from here."
- [6]
Vals AI observed: "The model appeared defeated, spending the next several hours doing essentially nothing but farming potatoes."
- [7]
Viewers watching the experiment live on Twitch started to agitate for the model to pick up the pace.
- [8]
Before the Creeper incident, GPT-6 Astra had spent well over 100 hours in the game.
- [9]
Tom's Hardware reports the model was recorded autonomously completing Portal in 24 hours at a cost of $571 in tokens.
- [10]
Astra had collected six blaze rods and three ender pearls; in the Nether it built a semi-automatic blaze farm, killed more than half a dozen Endermen in a warped forest, and gathered the ingredients for eyes of ender.
- [11]
Astra logged a note, partly in caps: "ALWAYS CARRY CRITICAL ITEMS with keepInventory; don't store in unguarded chest ever again."
- [12]
According to Vals AI, its Minecraft run works through general computer use, meaning screen, mouse and keyboard, with no specialized game hookup.
- [13]
The 2023 research project Voyager, which involved Nvidia and Caltech, had GPT-4 propose tasks in Minecraft, write JavaScript for them, and revise that code based on error messages; working programs went into a skill library.
- [14]
Voyager never saw the game: it received structured data and controlled Minecraft through the Mineflayer programming interface.
- [15]
ARC Prize explained that Astra translates unfamiliar game mechanics into compact symbolic descriptions and tracks objects, coordinates, rules and planned actions in a shorthand it develops itself.
- [16]
On ARC-AGI-3, Astra hit 62.7 percent over the standard interface and about 99.9 percent with OpenAI's own harness.
- [17]
On ARC-AGI-3, GPT-5.6 Sol scored 7.78 percent and Claude Opus 5 a little over 30 percent.
- [18]
According to Clad3815, operator of "GPT Plays Pokemon", Astra earned the champion title in Pokemon FireRed after 18 hours and 12 minutes.
- [19]
At the Portal run's hourly token cost, 141 hours would come to roughly $3,350.
- [20]
GPT-5.6 Sol needed 96 hours and 35 minutes to earn the champion title in Pokemon FireRed.
- [21]
GPT-5.5 still had not finished Pokemon FireRed after more than 218 hours.
- [22]
The user imjustnewatai played the Fallout 3 opening in Vault 101 himself and then handed the save file to OpenAI's agent tool Codex, with Astra set to maximum reasoning; from there the agent picked its own route and actions.
- [23]
In a community Factorio: Space Age run controlled through a custom Lua mod with an MCP interface, Astra produced blue science packs after about two hours and launched its first rocket after roughly ten.
- [24]
In the same Factorio setup, GPT-5.6 Luna and Fable 5.1 never got past power supply and oil prospecting.
- [25]
Neither the-decoder.com nor Tom's Hardware reports a token cost for the 141-hour Minecraft run.
- [26]
The gap between Astra's standard-interface ARC-AGI-3 score and its score under OpenAI's own harness is 37.2 percentage points.
- [27]
The Portal run cost about $23.80 per hour of wall-clock agent time.
- [28]
GPT-5.6 Sol took about 5.3 times as long as Astra to finish Pokemon FireRed.
- [29]
According to Tom's Hardware, OpenAI has claimed Astra is an ethereal 'Alien Mind' with AGI-like qualities.
Sources
2 independent publishers whose own reporting we read for this story.
- the-decoder.comGPT-6 Astra: Pokemon champion in 18 hours, potato farmer after one Creeper mishap
1 article · September 17, 2026
- tomshardware.com'Defeated' GPT-6 Astra model spent several hours just farming potatoes after being blown up by a Creeper in Minecraft — OpenAI offering gets further than any other AI system in 141-hour test
1 article · September 16, 2026
Topics and entities
Topics
- AI Agents
- Agent Benchmarks
- Video Games as AI Testbeds
- Inference cost
- Computer-Use Models