GPT-5.6 just tripled its score on one of AI's hardest benchmarks — not by retraining the model, but by flipping two API settings, and that distinction matters enormously for how we think about AI capability.
The GPT-5.6 Breakthrough on ARC-AGI-3
OpenAI has published a technical breakdown showing how two configuration changes — retaining reasoning state between turns and enabling context compaction — sent GPT-5.6's ARC-AGI-3 scores through the roof.
ARC-AGI-3 is widely regarded as one of the most demanding tests of general reasoning in AI, designed specifically to resist pattern-matching shortcuts. Tripling a score on it isn't a rounding error — it's a signal.
The key insight: the model wasn't getting smarter. It was being allowed to stay smart across a longer chain of thought without losing its thread or wasting compute re-establishing context.
What "Retained Reasoning" and "Compaction" Actually Do
Retained reasoning means the model's internal working memory persists between API calls, so it doesn't start from scratch every time. Think of it as the difference between solving a maths problem with a notepad versus having someone erase your working after every line.
Context compaction is the efficiency partner: instead of bloating the context window with raw token history, the model compresses prior reasoning into a dense summary. Less noise, more signal, lower cost per useful output.
Together, these settings let GPT-5.6 tackle multi-step problems the way a human expert would — building on prior conclusions rather than constantly re-deriving them. If you want to go deeper on why this architecture matters, our course on Future of AI Inference covers exactly how inference-time decisions reshape model performance.
What This Means for Learners
This story is a masterclass in a skill that's becoming essential: knowing how to configure AI, not just how to prompt it. The model capability was always there — the unlock was understanding the settings.
If you're building with AI APIs or managing AI tools at work, this is your reminder that default settings are rarely optimal settings. Digging into configuration is now a legitimate competitive advantage.
For a practical grounding in how these architectural choices play out in real deployments, GPT-5.6: The AI They Locked Down gives you the context to read OpenAI's technical releases with genuine fluency — not just headlines.
