GPT-6 Astra posts record agentic game results; OpenAI tells developers to lean out prompts and loosen guardrails

OpenAI's GPT-6 Astra beat GPT-5.6 Sol and Claude Opus 5 on agentic game benchmarks — finishing Pokemon FireRed in 18h 12m and scoring 62.7% on ARC-AGI-3 — while OpenAI's Eric Provencher published guidance advising developers to shorten skill descriptions,…

Two reports from The Decoder cover OpenAI's GPT-6 Astra. The first (2026-09-12) describes guidance from OpenAI's Eric Provencher on adapting developer setups for the new model: overly long or conflicting skill descriptions — skills being Markdown prompt files whose names and descriptions enter Codex's context — get truncated and cause wrong skill selection, so he recommends selective document references in AGENTS.md instead of full repository overviews before every change, explicit permissions for safe operations like throwaway-data local tests to cut repeated confirmation requests, and defining upfront what "done" means, since Astra may stop earlier than GPT-5.6 Sol even without restrictions. He also warns that shared skills affect all contributors' agents, so team-wide restrictions tuned for older models may be too strict. The second report (2026-09-17) details benchmark results: Astra completed Pokemon FireRed in 18h 12m versus 96h 35m for GPT-5.6 Sol and scored 62.7% on ARC-AGI-3 via the standard interface versus 7.78% for GPT-5.6 Sol and about 30% for Claude Opus 5. In a Vals AI Minecraft run driven through general computer use (screen, mouse, keyboard only, no specialized game hookup), the agent built a Nether portal within three hours, but the 141-hour run ended after a Creeper explosion triggered risk-averse potato farming for hours. ARC Prize attributes the leap to the model converting observations into compact, reusable symbolic rules it develops itself. Astra also completed runs of Portal, Fallout 2, Fallout 3, RimWorld, and Factorio: Space Age.

• ARC-AGI-3 score: GPT-6 Astra 62.7% (standard interface), GPT-5.6 Sol 7.78%, Claude Opus 5 about 30%
• Pokemon FireRed completion: GPT-6 Astra in 18h 12m versus 96h 35m for GPT-5.6 Sol
• Vals AI Minecraft run used only screen, mouse, and keyboard with no specialized game hookup; the agent built a Nether portal within three hours
• The 141-hour Minecraft run ended after a Creeper explosion; the agent wrote self-protective rules and farmed potatoes for hours
• ARC Prize attributes the performance leap to the agent turning observations into compact, reusable self-written symbolic rules
• GPT-6 Astra also completed runs of Portal, Fallout 2, Fallout 3, RimWorld, and Factorio: Space Age
• Skills are Markdown prompt files whose names and descriptions enter Codex's context; too many or conflicting skills cause truncation and wrong skill selection
• OpenAI's Eric Provencher recommends selective document reads in AGENTS.md, explicit permissions for safe operations like throwaway-data local tests, and defining upfront what "done" means