GPT-6 Astra: Pokémon Champion in 18 Hours, Factorio Rocket, Fallout 3 Completed

GPT-6 Astra records unprecedented performances in games and benchmarks while developing caution strategies after failures. Detailed sessions by ARC Prize and community operators show how the model transforms its observations into reusable rules, from speedrunning to potato monoculture.

Rules Taking Hold, Leading to Slowdown

In Fallout 3, after several deaths in Vault 87, GPT-6 Astra increased its saves and progress checks. It indicated that it checks each movement against the visible geometry to avoid any loss, a reflex that significantly slowed down the session without interrupting it. In Minecraft, right after a Creeper explosion, it imposed a strict rule on itself: keep important items on hand and avoid unprotected chests. Over time, it developed a distrust of any large green object, correcting a confusion with sugar cane, while being hard on itself and reminding that a mistake can cost items. It also refused to spend another night hunting "dark pink pixels" after a failure, retreating for several hours to a potato field according to Vals AI. This type of principle can become problematic and, in open worlds subject to randomness, lead to overcorrection and forgetting the objective.

ARC-AGI-3 Scores: 62.7% Standard, 99.9% with OpenAI Harness

On ARC-AGI-3, GPT-6 Astra achieved 62.7% on the standard interface and about 99.9% with the OpenAI harness. References mention 7.78% for GPT-5.6 Sol and just over 30% for Claude Opus 5. ARC Prize describes a mechanism where Astra converts unknown mechanics into symbolic descriptions, tracks objects, coordinates, rules, and plans in a shorthand language, and evolves observations into rules and then into plans. This tendency to quickly formalize experience can help understand new mechanics in just a few attempts and avoids endlessly repeating ineffective actions.

Minecraft: Portal in Under 3 Hours, Then Total Loss and Retreat

In Minecraft, Astra built a Nether portal in less than three hours, fought zombies, trapped a skeleton, and found its base. It constructed a semi-automatic blaze farm, obtained six blaze rods and three ender pearls in a warped forest, gathering components for ender eyes. Thousands of viewers witnessed these advancements before a Creeper explosion obliterated the chest and bed, resulting in the loss of items. Astra then noted a caution rule for itself and spent hours farming potatoes. The session, operated via screen, mouse, and keyboard without specialized API according to Vals AI, totaled well over 100 hours before being stopped at 141 hours. Vals AI claims that no AI had reached this level before. In comparison, OpenAI's VPT only achieved the diamond pickaxe in 2.5% of attempts, without getting close to a portal, and took 10 minutes to reach it in a separate comparison.

Pokémon FireRed: 18h 12m for the Title, Far Ahead of GPT-5.x

According to operator Clad3815, GPT-6 Astra claimed the champion title in Pokémon FireRed after 18 hours and 12 minutes. GPT-5.6 Sol required 96 hours and 35 minutes, and GPT-5.5 still had not finished after more than 218 hours.

Factorio: Blue Packs in 2 Hours, First Rocket in 10 Hours

During a community session of Factorio: Space Age, Astra was controlled using a custom Lua mod and an MCP interface. After about two hours, it generated blue science packs, and the first rocket was launched around the tenth hour. With the same setup, GPT-5.6 Luna and Fable 5.1 did not surpass the energy and oil prospecting stage.

Fallout 3: 59 Hours to Credits, Driven by Observe-Act Loops

Astra's Fallout 3 session, published by imjustnewatai, started after a manually played introduction in Vault 101, with the save then handed over to the Codex agent tool with maximum reasoning settings. The agent chose its route through Galaxy News Radio, searching for the father, the G.E.C.K., escaping Raven Rock, and Project Purity, alternating pauses, observation, and action, with saves, companions, and fast travel allowed. The credits rolled after about 59 hours. In Vault 87, after several deaths and ammunition shortages, Astra systematically diagnosed action failures, noting for example the delayed functioning of a bed (health 260/260) or an obstacle blocking the line of fire, and established a principle: check that an input worked before repeating it.

What Astra Does Alone, Voyager Equipped It

The Voyager project, involving Nvidia and Caltech in 2023, enabled GPT-4 to propose Minecraft tasks, write JavaScript code, and revise it according to errors, capitalizing programs into a skill library. However, Voyager did not perceive the game directly and relied on structured data and the Mineflayer API. While previous systems surrounded themselves with specialized tools, the approach described by ARC Prize for Astra consists of transforming its observations into rules and then into plans on its own.