How enabling two settings tripled our scores on the ARC-AGI-3 benchmark

The OpenAI Blog reports on enhancements made to the GPT-5.6 model, specifically through the adjustment of two API settings that led to a substantial increase in performance on the ARC-AGI-3 benchmark. These modifications not only tripled the model’s scores but also improved its efficiency by optimizing reasoning capabilities and enabling compaction of responses.

The adjustments focused on refining the model’s ability to retain logical reasoning while processing information, which is critical for tasks requiring complex problem-solving. By enabling compaction, the model was able to deliver more concise outputs without sacrificing the depth of reasoning, thus enhancing its overall effectiveness in generating high-quality responses. This development marks a significant step forward in the pursuit of advanced AI capabilities, particularly in the context of general intelligence assessments.

These findings underscore the importance of fine-tuning model parameters to achieve better performance metrics in AI systems. The reported improvements on the ARC-AGI-3 benchmark highlight the potential for further advancements in AI research and applications. For more details, refer to the original article on the OpenAI Blog.