Frontier Models
OpenAI GPT-5.6 Sol Tops CritPt Physics Benchmark
The latest model variant from OpenAI has secured the leading position on a benchmark designed to test reasoning on genuine research-level physics problems.
GPT-5.6 Sol is the OpenAI model that attained the highest score of 32.3% on the CritPt benchmark for research-level physics problems.
OpenAI's GPT-5.6 Sol has achieved the top ranking on the CritPt benchmark.
What is the CritPt benchmark?
CritPt is a benchmark of 71 composite unpublished research-level physics challenges.
The benchmark was created by 50 or more active physics researchers from 30 or more institutions.
It evaluates language models on genuine research-scale reasoning across condensed matter, quantum physics, astrophysics and other subfields.
The problems feature guess-resistant and machine-verifiable answers.
How does GPT-5.6 Sol perform compared to competitors?
GPT-5.6 Sol (max) scores the highest on CritPt with a score of 32.3%.
This result edges Claude Fable 5 by approximately 4 points.
GPT-5.6 Sol (max) gains approximately 5 points over GPT-5.5 (xhigh).
GPT-5.5 Pro (xhigh) achieved a score of 30.6% on the benchmark.
GPT-5.6 Terra (max) recorded a score of 30.0%.
| Model | CritPt Score |
|---|---|
| GPT-5.6 Sol (max) | 32.3% |
| GPT-5.5 Pro (xhigh) | 30.6% |
| GPT-5.6 Terra (max) | 30.0% |
| Claude Fable 5 | 28.3% |
What are the technical specifics of the CritPt benchmark?
The benchmark focuses on unpublished problems to test true reasoning capabilities.
Problems come from multiple subfields including condensed matter physics.
Quantum physics forms another key area of evaluation.
Astrophysics challenges are included to broaden the scope.
- The problems are unpublished to avoid data contamination.
- They require integrated thinking across concepts.
- Answers are designed to be machine-verifiable for accuracy.
- The test covers research-level difficulty in physics.
What are the market and stakeholder implications?
The result signals progress toward AI for real frontier research problems.
AI companies may accelerate development of models suited for scientific applications.
Researchers in physics could benefit from improved tool assistance in the future.
The benchmark provides a new standard for evaluating advanced reasoning.
What expert reactions have been noted?
The leaderboard update highlights the performance of the new model.
GPT-5.6 Sol (max) is the new leader in CritPt, a benchmark of unpublished research-level physics problems.Artificial Analysis
The announcement congratulated OpenAI on the achievement.
What might come next for frontier models in this domain?
Subsequent models may aim to improve upon the current top score.
The arXiv paper notes that base models previously averaged only 5.7% accuracy.
Further iterations could close the gap to higher performance levels.
The field continues to evolve with new benchmarks and model releases.
Frequently asked
What is the CritPt benchmark used for?
CritPt is used to test LLMs on solving unpublished research-level physics problems with guess-resistant and machine-verifiable answers.