📰 Key Highlights

OpenAI discovered that tweaking just two API settings can dramatically boost GPT-5.6’s scores on the ARC-AGI-3 benchmark while also improving efficiency. The first setting is retaining reasoning, which lets the model carry forward its reasoning chain across multi-turn interactions and multi-step tasks instead of rethinking from scratch every time — preserving logical coherence and cutting down on redundant computation. The second setting is enabling compaction, which compresses prior reasoning before passing it to subsequent steps, slashing context length and computational load without sacrificing key information. Combined, these two settings nearly tripled GPT-5.6’s scores on ARC-AGI-3 style tests that demand long-chain reasoning and multi-step problem solving, and because they avoid unnecessary re-reasoning and bloated context, overall computational efficiency improved too. This shows that for complex reasoning tasks, how you manage a model’s “memory” and “continuity of thought” is itself a direct lever for optimizing performance and cost — not something only fixable by swapping models or throwing more compute at it. Full technical implementation details and complete test data are in the original link.

💬 JudyAI Lab Perspective

Based on the original summary, no extra facts added, adopting JudyAI Lab’s observer viewpoint:

OpenAI’s latest tests show that adjusting two API settings can dramatically boost GPT-5.6’s scores on the ARC-AGI-3 benchmark — a low-cost optimization path worth every AI builder’s attention.

The two settings are retaining reasoning and enabling compaction: the former lets the model carry its reasoning chain forward across multi-turn tasks instead of rethinking from scratch; the latter compresses prior reasoning before handing it off to the next step, cutting context length and compute without losing key info. Combined, they nearly tripled GPT-5.6’s scores on multi-step tests requiring long-chain reasoning, with efficiency improving in parallel. This case highlights an often-overlooked angle: the bottleneck for complex reasoning tasks isn’t always the model’s raw capability. How you manage “memory” and “continuity of thought” is itself a direct lever for optimizing performance and cost — you don’t always have to swap models or scale up compute.

Next time your multi-step reasoning task underperforms, check whether your API calls are preserving reasoning state before you rush to upgrade the model.

📅 Source Info

🔗 Further Reading