... like button. Benchmarks, benchmarks, benchmarks Benchmarks should always be taken with a grain of salt . They are effectively a snapshot of a models capabilities under near perfect conditions. Sort of like the Big Mac you see on the ads vs the Big Mac you get in the bag. It gives you a good idea, but they're far from a perfect measure of real world usage. Math (AIME 2025, no tools): 94.6 percent Real-world coding (SWE-bench Verified ... ... Multilingual programming (Aider Polyglot): 88.0 percent Multimodal understanding (MMMU): 84.2 percent Medical reasoning (HealthBench Hard): 46.2 percent Graduate-level logic (GPQA without tools, via GPT-Thinking Pro): 88.4 percent In production, GPT-5 Thinking cuts hallucinations by 45 percent versus