OpenAI Reveals How Two API Settings Tripled GPT-5.6's ARC-AGI-3 Scores

OpenAI published a technical blog post today explaining how two API configuration settings tripled GPT-5.6's performance on the ARC-AGI-3 benchmark. The two settings — one for retaining reasoning capabilities and one for enabling model compaction — work in tandem to boost efficiency while preserving reasoning quality.

ARC-AGI-3 is a benchmark designed to test abstract visual reasoning, requiring models to identify underlying patterns in unfamiliar visual inputs. It has long been considered a rigorous test of a large language model's reasoning limits, making GPT-5.6's threefold score increase a notable milestone.

According to OpenAI, the reasoning-retention setting ensures that key chain-of-thought reasoning steps are preserved during model compression, while the compaction setting reduces computational overhead. Together, they produce substantial gains in both accuracy and efficiency.

The significance of this finding lies in the fact that API-level configuration changes — rather than full model retraining or fine-tuning — can yield such dramatic benchmark improvements. This offers practical insights for enterprises deploying large models in production environments.

The post also discusses how the two settings perform across different reasoning task types and provides guidance for developers looking to adapt the configurations to their own use cases. OpenAI encourages developers to experiment with these settings to explore further applications.

This development also raises questions about the future of the ARC-AGI benchmark series. If simple API parameter adjustments can produce such outsized gains, the benchmark's difficulty ceiling and discriminative power may need re-evaluation — a topic that will likely draw continued industry attention.