GPT-5.6 Sol Wins ARC-AGI-3 Game at 87%, Averaging 13.33% on Public Tests
Updated
Updated · arcprize.org · Jul 9
GPT-5.6 Sol Wins ARC-AGI-3 Game at 87%, Averaging 13.33% on Public Tests
2 articles · Updated · arcprize.org · Jul 9
Summary
GPT-5.6 Sol became the first model to win an ARC-AGI-3 public game, scoring 87% on ft09 while averaging 13.33% across the public benchmark and 7.78% on semi-private tasks.
The report attributes that lead to orientation rather than execution: Sol reads unfamiliar scenes in the game’s own vocabulary and replans after failed hypotheses instead of thrashing.
Verified leaderboard results show a sharp gap inside the GPT-5.6 family, with Sol Max posting 7.8% on ARC-AGI-3 versus 0.8% for Terra Max and 0.2% for Luna Max.
That contrast is narrower on older benchmarks, where Sol Max reached 96.5% on ARC-AGI-1 and 92.5% on ARC-AGI-2, suggesting ARC-AGI-3 remains the key bottleneck for frontier models.
GPT-5.6 mastered an abstract game. When will it solve real-world problems like climate change or disease?
As AI’s reasoning power skyrockets, is its immense environmental cost the biggest barrier to progress?
With AI now able to coordinate sub-agents, what new safeguards are needed to manage its emergent capabilities?
GPT-5.6 Sol’s 7.8% ARC-AGI-3 Breakthrough: Closing the Gap Toward Human-Level Generalization
Overview
OpenAI's GPT-5.6 Sol reached a historic milestone on July 9, 2026, by setting a new state-of-the-art performance on the challenging ARC-AGI-3 benchmark. Its 'Max' variant achieved a 7.8% score, making it the first verified frontier model to successfully complete an ARC-AGI-3 game. This breakthrough highlights Sol's exceptional ability to orient itself in novel situations it has never encountered before, serving as a key indicator of generalized intelligence. The achievement marks significant progress in AI's capacity for true generalization and problem-solving in unfamiliar environments.