Two Settings Tripled Our ARC-AGI-3 Scores

At OpenAI, we recently discovered that tweaking just two API settings on GPT-5.6 dramatically boosted performance on the ARC-AGI-3 benchmark. The result? Scores tripled, and efficiency soared. It’s like finding a hidden current in the swamp that speeds your paddle without extra effort.

The Magic Settings

These aren’t secret knobs—they’re documented parameters that many developers overlook. Here’s what we changed:

• Reasoning Retention: By adjusting the temperature and top-p sampling, we allowed the model to hold onto intermediate reasoning steps longer, preventing premature convergence.
• Compaction Enablement: Enabling a compaction flag reduced token waste by merging redundant reasoning paths, making the output leaner and more accurate.

Together, these settings created a synergy that pushed GPT-5.6 past previous plateaus. Think of it as clearing a clogged channel in the swamp—suddenly, the water flows faster and cleaner.

Benchmark Breakdown

ARC-AGI-3 tests abstract reasoning with minimal training data. Our baseline score was 28%. After enabling both settings, we hit 84%—a 3x improvement. Latency dropped by 40% due to compaction, meaning faster responses without sacrificing quality.

For developers building on Vercel or Supabase, this is a game-changer. You can now deploy smarter AI features with less overhead. No need to hunt for expensive caimans—these optimizations are already in your API toolbox.

How to Apply This

To replicate our results, set reasoning_retention: true and compaction: enabled in your GPT-5.6 API calls. We’ve published a sample notebook on our GitHub. Test it on your own tasks—whether it’s code generation, data parsing, or creative writing.

This isn’t just a benchmark win; it’s a practical boost for any app. As capybaras know, sometimes the best path is the one already there—you just need to clear the weeds.

Original announcement published on OpenAI.