Editorial illustration for GPT-6 Astra Trains to Play Pokemon in 18 Hours
GPT-6 Astra Beats Pokemon in 18 Hours, Smashes AI Benchmarks
GPT-6 Astra Trains to Play Pokemon in 18 Hours
Vals AI clocked GPT-6 Astra finishing Pokemon in 18 hours, a game that used to take 96. The same model launched a rocket in Factorio and rolled credits on Fallout 3, clearing benchmarks that stopped every earlier system cold. Then it went into Minecraft and spent hours on potatoes.
The run itself was already a record. No AI system had gotten as far in Minecraft as Astra did, according to Vals AI, which tracked the model through more than 140 hours of play. It built a blaze farm in the Nether, took down Endermen in a warped forest, and assembled the eyes of ender needed to find the game's final portal. People watched live as Astra loaded its rewards into a chest, work that represented dozens of hours of progress toward beating the game outright.
What happened next is the part worth pulling out on its own. A single Creeper, a single explosion, and a model that had just out-performed every predecessor on record found itself starting over from close to nothing, forced to rebuild with whatever it could scrounge from the dirt.
Pokemon champion in 18 hours instead of 96, a rocket launch in Factorio, the credits rolling in Fallout 3. GPT-6 Astra beats games that stumped every model before it. But then it spent hours farming potatoes in Minecraft after a Creeper blew up its stash.
Why this matters
The Pokemon and Minecraft numbers are real progress, and worth tracking if you're benchmarking agentic reasoning: an 18-hour FireRed clear against GPT-5.6 Sol's 96 hours is a five-fold jump in one model generation, and getting further in Minecraft than VPT or any prior system tells us something about long-horizon planning under partial information. But the potato-farming detour after a Creeper wipes out its supplies is the more useful data point for anyone building on this model. It shows the gap between "can execute a long plan" and "can recover gracefully when the plan breaks." For developers wiring Astra into agents that touch real infrastructure or money, that's the failure mode to stress-test, not the speedrun highlight.
Vals AI and Clad3815's tracker are doing the field a favor by publishing granular, game-by-game numbers instead of one composite score. We'd rather see ten narrow benchmarks like this than another leaderboard. Watch what Astra does the next time something blows up its stash, not just how fast it beats the game.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives OpenAI’s GPT-6 Astra Decrypts Unsolved Nazi Enigma Message in Ten Hours
From Archives Journalist Questions AI Existential Threat Scenarios
Common Questions Answered
How much faster did GPT-6 Astra complete Pokemon FireRed compared to GPT-5.6 Sol?
GPT-6 Astra completed Pokemon FireRed in 18 hours, compared to GPT-5.6 Sol's 96 hours, representing a five-fold improvement in speed between model generations. This significant reduction in completion time demonstrates substantial progress in the model's ability to play complex games efficiently.
What games did GPT-6 Astra successfully complete according to the benchmarks mentioned?
GPT-6 Astra successfully completed Pokemon FireRed, launched a rocket in Factorio, and rolled credits on Fallout 3. These achievements represent breakthroughs as each of these games had previously stopped every earlier AI system from completing them.
What was GPT-6 Astra's record-breaking achievement in Minecraft?
GPT-6 Astra achieved the furthest progress in Minecraft of any AI system to date, according to Vals AI, which tracked the model through more than 140 hours of gameplay. During this extended play session, the model built a blaze farm in the Nether, demonstrating advanced long-horizon planning capabilities.
What does the potato farming incident in Minecraft reveal about GPT-6 Astra's behavior?
After a Creeper explosion destroyed its supplies in Minecraft, GPT-6 Astra spent hours farming potatoes instead of pursuing other objectives. This detour is described as a more useful data point for understanding how the model adapts and makes decisions under partial information and unexpected setbacks.
Why is the improvement in agentic reasoning important for benchmarking GPT-6 Astra?
The five-fold speed improvement in Pokemon and the unprecedented progress in Minecraft demonstrate significant advances in long-horizon planning under partial information, which are key metrics for evaluating agentic reasoning capabilities. These benchmarks show that GPT-6 Astra can handle complex, multi-step tasks that require sustained decision-making over extended periods.
Further Reading
- [AINews] GPT-6 Astra: OpenAI's biggest LLM launch of all - Latent Space
- GPT-6 Astra - Vals AI
- GPT-6 Astra : les évaluateurs indépendants tempèrent les affirmations d’OpenAI - LeMagIT
