Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
external benchmark
End-to-End Physics Research Benchmark for AI Agents. Every model the catalog carries a published FrontierPhysics value for, ranked by that value.
FrontierPhysics, from BenchFlow, gives an AI agent a research project written by a working physicist and checks whether it gets the physics right and writes it up well enough to publish. It has 56 tasks across nine fields of physics, and each one is run three times. 21 models in the ModelScale catalog have a published value.
Each point is a model with a published FrontierPhysics value and complete token prices, on the 75% input / 25% output blend. Frontier models are spaced evenly along the cost axis.
21 models with a published FrontierPhysics value, ordered by that value, highest first. Models the source has not scored on this benchmark are not listed — they are unmeasured, not last.
| Rank | Model | Provider | Agentic research in a sandbox: literature review, simulation or measurement, structured results and a written paper |
|---|---|---|---|
| 1 | Anthropic | 27.4 | |
| 2 | OpenAI | 23.2 | |
| 3 | OpenAI | 19.6 |
What the benchmark measures and how its value is published, from the source’s own record.
FrontierPhysics (End-to-End Physics Research Benchmark for AI Agents) is an external benchmark published by FrontierPhysics: Evaluating Agents for End-to-End Physics Research (arXiv, November 4, 2026). FrontierPhysics, from BenchFlow, gives an AI agent a research project written by a working physicist and checks whether it gets the physics right and writes it up well enough to publish. It has 56 tasks across nine fields of physics, and each one is run three times. Its evaluation set is described by the source as 56 research projects from 42 physicists across nine fields, three trials each (168 rollouts per setup). The source rates its difficulty as open-ended research projects that took their authors about 270 days on average.
Each model's result is published by FrontierPhysics: Evaluating Agents for End-to-End Physics Research (arXiv, November 4, 2026) as agentic research in a sandbox: literature review, simulation or measurement, structured results and a written paper. ModelScale prints that value to 1 decimal place, exactly as published, and never converts or rescales it. This page ranks the 21 models that carry a published value, highest value first; a model the source never scored is left out rather than ranked at zero. The source does not state whether a higher value is the better result, so neither does this page.
| 4 | OpenAI | 14.3 |
| 4 | OpenAI | 14.3 |
| 6 | Anthropic | 11.3 |
| 7 | Anthropic | 9.5 |
| 8 | 8.3 |
| 9 | OpenAI | 7.1 |
| 10 | Moonshot AI | 6.5 |
| 11 | Anthropic | 4.2 |
| 11 | Z.AI | 4.2 |
| 11 | xAI | 4.2 |
| 14 | DeepSeek | 3.6 |
| 14 | OpenAI | 3.6 |
| 14 | OpenAI | 3.6 |
| 17 | xAI | 2.4 |
| 17 | Moonshot AI | 2.4 |
| 19 | Xiaomi | 1.2 |
| 20 | 0.0 |
| 20 | Mistral | 0.0 |
Evidence key: Observed
Rows are ordered by the value FrontierPhysics: Evaluating Agents for End-to-End Physics Research (arXiv, November 4, 2026) published, highest first. The source does not state whether a higher value is the better result, so this page does not either: for a benchmark that measures a rate of failure, read the table from the bottom.