DeepSeek
DeepSeek V4 Pro
Release date: 4/23/2026
A cost-efficient open reasoning model for mathematics, code, and agentic workflows.
Pramaana IndexAccuracy: StandardLatency: StrongCost: Leading
- Accuracy / Pramaana Index
- 62.48%± 0.82
- Latency / Pramaana Index
- 344.30s
- Cost / Test / Pramaana Index
- $0.24
- Context window
- 256k
- Max output tokens
- 64k
- Input modality
- Text
Hyperparameter settings
temperature: 0.2
reasoning: extended
fallbacks: disabled
Benchmark performance for DeepSeek V4 Pro
| Benchmarks | Accuracy | Ranking |
|---|---|---|
| Pramaana Index | 62.48% ± 0.82 | 11 / 35 |
| Pramaana Multimodal Index | 60.47% ± 0.99 | 11 / 24 |
| Legal Research Fidelity | 56.88% ± 1.16 | 11 / 18 |
| Financial Modeling Agent | 52.59% ± 1.33 | 11 / 22 |
| Clinical Documentation Safety | 56.18% ± 1.50 | 11 / 19 |
| MATH-500 | 74.98% ± 1.67 | 7 / 112 |
| GPQA Diamond | 74.29% ± 0.48 | 11 / 121 |
| MMLU-Pro | 71.78% ± 0.65 | 11 / 119 |
| Adaptive Tutoring Dialogue | 58.08% ± 0.82 | 11 / 16 |
| SWE-bench Verified | 54.09% ± 0.99 | 9 / 69 |
| LiveCodeBench | 73.07% ± 1.16 | 5 / 127 |
| Agentic Reliability | 48.67% ± 1.33 | 11 / 12 |
| Strategic Games Arena | 54.69% ± 1.50 | 11 / 20 |
| Benefits Navigator | 57.99% ± 1.67 | 11 / 16 |