OpenAI's GPT-6 Astra Surpasses Human Performance on ARC-AGI-3
The AI completed all human-solvable games with fewer moves than median human players.
1GPT-6 Astra's Superior Performance
OpenAI's GPT-6 Astra model achieved 62.7% and 99.9% success rates on the ARC-AGI-3 benchmark using two different testing methods. With the Provider Adapter approach, it surpassed human performance at a cost of $19,000. ARC-AGI-3 is designed to measure AI capabilities in exploration, goal inference, and planning within novel environments with minimal prior knowledge.
2Revolution in Action Efficiency
Astra completed 96% of tested levels using fewer actions than human median. On average, it required only 51.7% of human moves to solve levels. These results demonstrate AI has reached human-level not just in problem-solving ability but also in solution efficiency, marking a significant milestone in agentic intelligence development.
3Symbolic Modeling Capability
GPT-6 Astra demonstrated remarkable ability to transform unfamiliar environments into compact symbolic models. It tracked object positions, interaction rules, and required action sequences using a code-like notation. This capability allows the model to quickly comprehend complex problems and develop effective solution strategies.
4The ARC-AGI Benchmark Series
ARC-AGI-3 is the third-generation platform designed to measure progress toward artificial general intelligence. It features more complex scenarios and less prior knowledge than previous versions. The environments are 100% solvable by humans, with difficulty calibrated through controlled testing. Each new generation evolves alongside advancements in AI capabilities.
5Cost Comparison Analysis
While human participants earned approximately $12.78 per attempted game, Astra's solution cost ranged between $19,000-$26,000. However, when considering only brain energy as electricity cost, human performance amounts to about 0.067 cents per game. This comparison highlights that AI systems still lag behind human brain energy efficiency.
6Custom Notation System
Astra developed its own domain-specific language notation to track game mechanics. This system compactly represented object coordinates, interaction rules, and incomplete plans. Although not a full programming language, this algebraic shorthand enabled rapid environment modeling and efficient solution planning.
7Human-AI Performance Gap
Pre-launch predictions suggested AI would require substantially more attempts than humans, even when solving problems. Contrary to brute-force approaches, Astra operated within human efficiency ranges once it 'understood' game mechanics. This finding indicates advanced AI models have made qualitative leaps in cognitive efficiency.
8Significant Step Toward AGI
GPT-6 Astra's performance on ARC-AGI-3 demonstrates substantial progress in human-like problem-solving and adaptation. Its action efficiency and symbolic modeling skills represent critical milestones toward AGI. These advancements suggest AI's growing potential to operate effectively in increasingly complex and uncertain real-world environments.