GPT-5.6 Sol (max) Tops CritPt, a New Benchmark of Unpublished Physics Research Problems
GPT-5.6 Sol (max) leads CritPt, a new physics benchmark built from unpublished graduate-level research problems, scoring about 5 points above GPT-5.5 and 4 points ahead of Claude Fable 5, according to independent testing by Artificial Analysis.
CritPt was developed by Argonne National Laboratory and the University of Illinois Urbana-Champaign specifically to sidestep a growing problem with AI physics benchmarks: once a problem and its answer are public, models can absorb it during training and appear to "reason" their way to a solution they actually memorized. CritPt instead draws on unpublished, graduate-level research problems contributed by more than 60 researchers across over 30 institutions worldwide, giving models no way to have seen the answers in advance.
On Artificial Analysis's independent scoring, GPT-5.6 Sol (max) leads the field with roughly a 32% score, gaining about 5 points over its predecessor, GPT-5.5 (xhigh) (GPT-5.6 just made every dollar in AI count harder), and finishing roughly 4 points ahead of Anthropic's Claude Fable 5 (Claude Fable 5 and Mythos 5: The Future of AI Is Gated…). The rest of the field drops off sharply from there, with most other frontier models scoring in the single digits to low twenties.
Even at the top of the leaderboard, a score near 32% is a reminder of how hard the benchmark is by design: these are problems drawn from active physics research, not textbook exercises, and no model is getting close to solving most of them. Artificial Analysis framed the result as an early signal of frontier models' potential usefulness in real scientific research rather than evidence that any model can meaningfully assist with it yet.