How GPT-6 Astra solved ARC-AGI-3: an interactive move-by-move breakdown

GregKamradt · x · 2026-10-01

MTSlive published an interactive deep-dive on ARC-AGI-3, amplified by Greg Kamradt.

• ARC-AGI-3, built by the ARC Prize Foundation, asks models to enter unfamiliar games, discover their rules, and learn to finish efficiently.
• The piece lets readers first play one public environment (SK48) themselves before unpacking what the score means.
• In ARC Prize's open-source harness, each turn arrives as plain text: game state, a 64×64 grid of integers 0–15, and legal actions, with a single system message: "You are playing a game. Your goal is to win."
• On Aug 31, 2026, ARC Prize recorded GPT-6 Astra at max reasoning effort clearing all 8 levels in 242 moves over 90 minutes, with Astra's per-move notes reproduced verbatim.
• The replay shows Astra forming a rules theory on move 1 (dots must reach crosses, button boxes drive the bars) and testing it with a single click — naming objects "orange" and "blue" despite never seeing color.

Kamradt noted the benchmark ecosystem is becoming an increasingly important measure of progress and capability transparency.

Original post →

More from Models

• Interleaved Head Attention Accepted at NeurIPS, Boosts RULER Retrieval 10-20% — _arohan_ · 2026-10-01
• Dev praises Opus 5.5 status messages for reporting progress like an expert engineer — sytelus · 2026-10-01
• Untrained d1 model plays Smash Melee at impressive level, researcher stunned — maximelabonne · 2026-10-01
• GPT-6 Astra prefers CDT over EDT, unlike Anthropic models, suggests decision-theory sycophancy gap — dfrsrchtwts · 2026-10-01
• OpenAI makes Luna free for subscribers; dev says gap to bigger models is now narrow — dosco · 2026-10-01
• 8B model trained on distributed gaming GPUs for $6,500 runs on phone CPU at ~60 tok/s — markjeffrey · 2026-10-01

Models channel →