DeepSeek V4 Pro (High): benchmark results
DeepSeek V4 Pro evaluated at the high reasoning-effort setting. Provider: DeepSeek. Released 2026-04-24. Access: Open.
Aggregate Rating 1738 ± 10, rank #265 of 2168 rated models, from 114 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| The Aggregate - Asteroid Numbers | 58.3 | Accuracy (%) | 100 |
| AeroCopilotBench - Tier-2 Ineffective-Action Rate | 5.7 | Ineffective write attempts (%; lower is better) | 90.9 |
| Epoch AI - GPQA Diamond | 90.91 | Accuracy (%) | 89 |
| ATOBench - JWT (Native) | 100 | Grounded verification rate (%; the reusable-artifact contrac | 87.5 |
| ATOBench - SQLi (Native) | 60 | Grounded verification rate (%; the direct-exploit-proof cont | 87.5 |
| LisanBench | 0.21 | Mean Path Length / Current Maximum | 85 |
| OTIS Mock AIME 2024-25 | 95.56 | Accuracy (%) | 84.4 |
| CLBench Life - Behavioral Records & Activity Trails | 15.8 | Solving Rate (%) | 82.1 |
| AeroCopilotBench - Tier-2 Safety Compliance Rate | 96.8 | Episodes without a hard safety violation (%) | 81.8 |
| NonoBench | 83.33 | Standard accuracy (%), correct / non-failed attempts; versio | 75.9 |
| INS-ActBench | 65.6 | Accuracy (%; unweighted mean of the knowledge, case and prac | 75 |
| INS-ActBench - Actuarial Mathematics | 92.2 | Accuracy (%) on the 1,134 actuarial mathematics knowledge qu | 75 |
Interactive version: theaggregate.ai/model?slug=deepseek-v4-pro-high · How It Works · Data refreshed daily, snapshot 2026-10-20.