CausalVerify: leaderboard
CausalVerify (Zhang et al., 2026) tests whether language models can carry out econometric causal-estimation workflows that produce the right number, not just code that runs. For 100 synthetic difference-in-differences, event-study, instrumental-variables and regression-discontinuity scenarios, the model writes R code from a research question, a data description and a preview of the data, the code is executed, and the treatment effect it prints must fall within 50 percent relative error of a canonical estimator run on the same data; pass rate in percent with one attempt per scenario, higher is better.
Metric: Execution-grounded pass rate (%; L2b+ over 100 scenarios: 30 difference-in-differences, 24 event-study, 24 instrumental-variables and 22 regression-discontinuity; the model writes R code for a fixed-seed synthetic scenario from the research question, data description and a preview of the realised CSV; the code is executed and passes when the treatment-effect estimate it reports (read by a Claude Haiku 4.5 coefficient-extraction judge, event-study windows mapped to the canonical scale) is within 50% relative error of a canonical estimator applied to the same realised data; single-shot). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 88 | #100 |
| 2 | GPT-5 | 72 | #163 |
| 3 | GPT-4o | 62 | #741 |
| 4 | Claude Sonnet 4 (20250514) | 50 | #386 |
| 5 | O3 | 46 | #224 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=causalverify · How It Works · Data refreshed daily, snapshot 2026-11-12.