End-to-End Physics Research Benchmark for AI Agents (FrontierPhysics)
We show this table for reference; we do not rank on it.
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.
Benchmark score on FrontierPhysics — November 13, 2026
We compile the FrontierPhysics rows from benchmark-owner or independent runs. Claude Opus 5.5 leads the table at 27.4%, followed by GPT-6 Astra (23.2%) and GPT-6.1 Sol (19.6%). We do not use these results to rank models overall.
Claude Opus 5.5
Anthropic
Run by FrontierPhysicsGPT-6 Astra
OpenAI
Run by FrontierPhysicsGPT-6.1 Sol
OpenAI
Run by FrontierPhysics21 modelsAgenticCurrentDisplay onlyUpdated November 13, 2026
Benchmark score table (21 models)
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5AnthropicRun by FrontierPhysics | 27.4% | Run by FrontierPhysics | Not reported | Closed |
| 2 | GPT-6 AstraOpenAIRun by FrontierPhysics | 23.2% | Run by FrontierPhysics | Not reported | Closed |
| 3 | GPT-6.1 SolOpenAIRun by FrontierPhysics | 19.6% | Run by FrontierPhysics | Not reported | Closed |
| 4 | GPT-6.1 TerraOpenAIRun by FrontierPhysics | 14.3% | Run by FrontierPhysics | Not reported | Closed |
| 5 | GPT-6 SolOpenAIRun by FrontierPhysics | 14.3% | Run by FrontierPhysics | Not reported | Closed |
| 6 | Claude Fable 5.1AnthropicRun by FrontierPhysics | 11.3% | Run by FrontierPhysics | Not reported | Closed |
| 7 | Claude Opus 5AnthropicRun by FrontierPhysics | 9.5% | Run by FrontierPhysics | Not reported | Closed |
| 8 | Gemini 4 ArgonGoogleRun by FrontierPhysics | 8.3% | Run by FrontierPhysics | Not reported | Closed |
| 9 | GPT-5.6 SolOpenAIRun by FrontierPhysics | 7.1% | Run by FrontierPhysics | Not reported | Closed |
| 10 | Kimi K3.1 ProMoonshot AIRun by FrontierPhysics | 6.5% | Run by FrontierPhysics | Not reported | Open |
| 11 | Claude Haiku 5.5AnthropicRun by FrontierPhysics | 4.2% | Run by FrontierPhysics | Not reported | Closed |
| 12 | GLM-5.3Z.AIRun by FrontierPhysics | 4.2% | Run by FrontierPhysics | Not reported | Open |
| 13 | Grok 4.7xAIRun by FrontierPhysics | 4.2% | Run by FrontierPhysics | Not reported | Closed |
| 14 | GPT-5.6 TerraOpenAIRun by FrontierPhysics | 3.6% | Run by FrontierPhysics | Not reported | Closed |
| 15 | GPT-5.6 LunaOpenAIRun by FrontierPhysics | 3.6% | Run by FrontierPhysics | Not reported | Closed |
| 16 | DeepSeek V4.1 Flash 1017DeepSeekRun by FrontierPhysics | 3.6% | Run by FrontierPhysics | Not reported | Open |
| 17 | Kimi K3Moonshot AIRun by FrontierPhysics | 2.4% | Run by FrontierPhysics | Not reported | Open |
| 18 | Grok 4.6xAIRun by FrontierPhysics | 2.4% | Run by FrontierPhysics | Not reported | Closed |
| 19 | MiMo-V2.6-ProXiaomiRun by FrontierPhysics | 1.2% | Run by FrontierPhysics | Not reported | Open |
| 20 | Gemini 3.8 FlashGoogleRun by FrontierPhysics | 0.0% | Run by FrontierPhysics | Not reported | Closed |
| 21 | Mistral Large 4MistralRun by FrontierPhysics | 0.0% | Run by FrontierPhysics | Not reported | Open |
- Lab-reported
- Benchmark-run
- Independent
How to read this leaderboard
Higher is better, and every score is low. Claude Opus 5.5 in Claude Code leads at 27.4%, passing 46 of its 168 rollouts, ahead of GPT-6 Astra at max effort (23.2%) and GPT-6.1 Sol (19.6%), though the paper finds the top four setups cannot be separated statistically. The verifier pass rate is much higher (45.8% for Opus 5.5): many agents get the numbers right but fail the rubric because the research is not rigorous enough. Gemini 4 Argon shows the gap most clearly, passing the verifier on 31.5% of rollouts but both checks on only 8.3%.
Operator receipt: 21 sourced rows are currently displayable on this page; the leading published row is Claude Opus 5.5 at 27.4%.
Honest limit: Each score belongs to a model and its agent tool together, and labs without their own tool run in Claude Code, so the table compares setups as much as models. With 56 tasks and three trials the intervals are wide: the 95% interval for Claude Opus 5.5 runs from 17.9% to 36.9%, and the paper finds GPT-6 Astra at either effort and GPT-6.1 Sol cannot be separated from it. The paper grade comes from one AI reviewer; the authors checked it against an expert audit and found that a Claude Opus 5.5 reviewer was more lenient on rollouts from its own model family. The leaderboard page states no license; the paper, which reports the same results, is CC BY 4.0. Where a model was run at two efforts we store the max run and name the other in the note, and we leave out DeepSeek V4 Pro because the page does not say which checkpoint it ran. We show it as display-only; it does not enter overall or category rankings.
Among the reported FrontierPhysics rows, Claude Opus 5.5 is first at 27.4%. The third row is 7.8 points behind. The broader top-10 range is 20.9 points, so the table still separates the published systems.
21 models have been evaluated on FrontierPhysics. The benchmark falls in the Agentic category. We keep these results separate from the weighted global scoring system, so they remain source-specific evidence. FrontierPhysics is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About FrontierPhysics
Year
2026
Tasks
56 research projects from 42 physicists across nine fields, three trials each (168 rollouts per setup)
Format
Agentic research in a sandbox: literature review, simulation or measurement, structured results and a written paper
Difficulty
Open-ended research projects that took their authors about 270 days on average
Forty-two physicists from 26 institutions turned projects from their own research into tasks, and an admin team of physicists accepted 56 of 182 submissions. The tasks cover condensed matter, atomic and optical physics, quantum information, astrophysics, fluids, statistical physics, photonics, high energy and chemical physics, and mix theory, simulation and experiment; some give the agent a simulated lab device to measure. The agent reads a proposal, searches the literature, plans and runs the work, then hands in its numbers in a fixed format and a paper. A deterministic verifier checks the numbers against reference values, and a reviewer agent (GPT-5.6 Sol at xhigh effort in Codex) grades the paper, code and transcript against rubrics: global ones shared by every task, including checks for reward hacking and invented results, and task-specific ones the task author wrote. A rollout passes only if it passes every verifier test and every blocking rubric criterion. Each model runs in its lab's own agent tool where there is one (Claude Code, Codex, Antigravity or Grok Build); other labs' models run in Claude Code. We store the overall pass rate and keep the verifier pass rate, pass@3 and the paper's 0 to 100 Score (the rubric grade of rollouts that pass, zero otherwise) in each row note.
Freshness and provenance
Version
FrontierPhysics 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does FrontierPhysics measure?
Whether an AI agent can carry a real physics research project from start to finish: read a proposal, search the literature, run simulations or measurements, report exact results and write a paper. A rollout counts only if its numbers match the reference and an AI reviewer passes the paper against the task author's rubric.
Which model leads FrontierPhysics?
Claude Opus 5.5 in Claude Code leads at 27.4%, passing 46 of 168 rollouts. GPT-6 Astra at max effort in Codex is second at 23.2%, and GPT-6.1 Sol and GPT-6 Astra at xhigh effort tie at 19.6%; the paper finds these four cannot be separated statistically. The best model from a lab outside the US is Kimi K3.1 Pro at 6.5%.
Why are FrontierPhysics scores so low?
Passing needs both correct numbers and research a physicist would accept. Agents often get the numbers but fail the rubric: Claude Opus 5.5 passes the verifier on 45.8% of rollouts but both checks on 27.4%. The paper finds the main weaknesses are loose reasoning and weak expert intuition, not missing physics knowledge or broken code.
Compare top models on FrontierPhysics
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.