GPT-5.4 pro nears Gemini on ARC-AGI-2

More than a victory lap, this is a sign that the frontier reasoning race is compressing into tiny single-digit gaps on benchmarks that still feel meaningfully hard.

• –The headline number matters because ARC-AGI-2 is designed to test abstraction and adaptability, not just polished benchmark memorization.
• –GPT-5.4 pro's 83.3% puts OpenAI back in striking distance of Gemini 3.1 Pro instead of clearly trailing on fluid-reasoning optics.
• –The jump from 74.0% for base GPT-5.4 to 83.3% for pro shows how much performance is now coming from extra reasoning effort, not just the base model.
• –ARC Prize attached a $16.41 per-task figure to GPT-5.4 pro, so this score is impressive but not cheap.
• –Developers should read this as a strong research signal, not a universal winner badge; real coding, agent, and tool-use workloads still matter more than a single benchmark.