VTC-Bench (Code Tools) - Spatial Reasoning: leaderboard
Metric: Pass rate (%) on spatial reasoning under extreme conditions (45 questions), answering VTC-Bench's 680 visual questions (538 multiple-choice, 142 open-ended) that call for chains of image operations with 32 OpenCV-based tools invoked by writing code (Qwen-Agent, or the Thyme framework for the open tool-use models), up to 20 rounds; rule matching plus a GPT-4o judge for open answers; proprietary models in high reasoning mode; higher is better. Source: arxiv.org. Saturation forecast: Around July 2027. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (High) | 55.56 | #77 (Gemini 3 Pro) |
| 2 | GPT-4o | 51.11 | #333 |
| 3 | GPT-5.2 (High) | 51.11 | #105 (GPT-5.2) |
| 4 | O3 (High) | 51.11 | #121 (O3) |
| 5 | Gemini 3 Flash (High) | 51.11 | #93 (Gemini 3 Flash) |
| 6 | Qwen 3 VL 235B A22B Instruct | 46.67 | #264 |
| 7 | O4 Mini (High) | 46.67 | #172 (O4 Mini) |
| 8 | Qwen 3 VL 30B A3B (Thinking) | 40 | #338 (Qwen 3 VL 30B A3B) |
| 9 | Qwen 3 VL 30B A3B Instruct | 37.78 | #365 |
| 10 | Qwen 3 VL 235B A22B (Thinking) | 37.78 | #228 (Qwen 3 VL 235B A22B) |
| 11 | Qwen 3 VL 32B (Thinking) | 33.33 | #287 (Qwen 3 VL 32B) |
| 12 | Qwen 3 VL 32B Instruct | 31.11 | #276 |
| 13 | Qwen 3 VL 8B (Thinking) | 20 | |
| 14 | Qwen 3 VL 8B Instruct | 17.78 | #401 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=vtc-bench-code-tools-spatial-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-11.