ARC-AGI-3 is now saturated, GPT-6 Astra reaches 99% with a new provider-adapter harness and beats human performance on 96% of ARC-AGI-3 levels.

A benchmark designed to expose failures in novel reasoning is now close to a ceiling for the strongest system.

But the interesting part is that, the exact same model looks radically different depending on how the evaluation system manages its memory between actions.

With ARC Prize's provider-neutral Standard harness, GPT-6 Astra tops out at 62.7%. Give Astra OpenAI's native way of preserving reasoning state and managing long context, and it jumps to 99.9%. The model did not suddenly get new weights. The environment did not become easier. The evaluation wrapper changed.

So "Provider Adapter harness" basically means:

Run the ARC benchmark while letting the model use the context-management machinery its own provider designed for it.

That is different from ARC's Standard harness, which deliberately tries to provide a minimal, common interface that works similarly across different model providers. ARC describes these as answering 2 different questions: the Standard harness asks how models compare under the same provider-neutral setup, while the Provider Adapter asks how well a model performs when used with the context features its provider built for it.