Two Settings Tripled GPT-5.6 ARC-AGI-3 Scores

OpenAI researchers have uncovered a surprising finding: enabling two specific API settings can dramatically improve GPT-5.6's performance on the ARC-AGI-3 benchmark, effectively tripling its scores. The two settings involve retaining reasoning traces and enabling a compaction feature, which together boost both accuracy and efficiency. This discovery provides a valuable insight into how model capabilities can be unlocked through configuration rather than additional training. The ARC-AGI-3 benchmark is designed to test general intelligence and reasoning abilities, making it a challenging metric for AI models. The fact that simple configuration changes can yield such significant improvements is noteworthy. It suggests that current models may be underutilized, with their full potential only accessible through the right settings. For developers and enterprises, this finding offers a cost-effective way to enhance performance. Instead of investing in expensive retraining or larger models, users can achieve better results by adjusting their API parameters. This is particularly beneficial for organizations with limited budgets or those looking to optimize their existing AI infrastructure. The reasoning traces setting allows the model to retain its internal thought process, which can improve decision-making and problem-solving. The compaction feature, on the other hand, helps streamline the model's output, reducing redundancy and improving clarity. Together, these settings create a more efficient and accurate reasoning loop. OpenAI's research team is now exploring whether similar configuration tweaks can unlock performance gains in other models and benchmarks. This line of inquiry could lead to a new era of AI optimization, where the focus shifts from building bigger models to using existing ones more effectively. As the AI industry matures, such insights will be crucial for maximizing the value of AI investments and pushing the boundaries of what is possible with current technology.