“OpenAI used the ARC-AGI benchmark to demonstrate the new reasoning capabilities of its early models because it was an unsaturated benchmark that could show a clear performance difference.”