GPT-5.4 tops CritPt physics benchmark

This matters less as a victory lap for one model and more as evidence that frontier evals are finally getting closer to real research work. CritPt is useful precisely because it shows how far models still are from acting like dependable physics collaborators.

• –CritPt covers 71 composite challenges and 190 checkpoints across 11 physics subfields, making it much harder to game than standard math or coding leaderboards.
• –Artificial Analysis highlights a big gap between current LLM performance and research-grade reasoning, so even a chart-topping 20% should be read as early progress, not scientific automation.
• –The benchmark is more meaningful than generic reasoning tests because answers are machine-verifiable and built around unpublished problems, which reduces contamination risk.
• –For AI developers, the bigger signal is evaluation direction: labs are being forced toward domain-specific, harder-to-cheat benchmarks that better reflect practical scientific use cases.