Backtrader-Bench - Mined Questions (No Tools): leaderboard
Backtrader-Bench (2026) tests LLM coding agents on algorithmic-trading questions whose answers come from executing Backtrader backtests, as four-option multiple choice with chance 25 percent. Accuracy on 38 harder mined questions answered with tools disabled, in percent; higher is better.
Metric: Accuracy (%; one run on the 38 four-option questions mined by a GPT-5.5 generator with executable verification and kept only when a tool-free GPT-5.4 solver failed and a tool-using solver succeeded; answered through the Cursor SDK with every tool disabled; chance 25%). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 60.5 | #91 |
| 2 | GPT-5.5 | 60.5 | #48 |
| 3 | Claude Opus 4.6 | 57.9 | #100 |
| 4 | Claude Opus 4.7 | 52.6 | #75 |
| 5 | Claude Sonnet 4.6 | 39.5 | #158 |
| 6 | GPT-5.3 Codex | 34.2 | #114 |
| 7 | Gemini 2.5 Flash | 29 | #452 |
| 8 | Grok 4.3 | 21.1 | #332 |
| 9 | Kimi K2.5 | 15.8 | #252 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=backtrader-bench-mined-questions-no-tools · How It Works · Data refreshed daily, snapshot 2026-11-12.