Two sets of data: DeepSeek's officially published scores, and a third-party cross-comparison (Z.ai official blog, 2026-08-14). Both are reproducible from MIT open weights.
| Evaluations | Models | Results | Notes |
|---|---|---|---|
| DeepSWE v1.1 (Pass@1) | V4.1-Flash-1017 | 77.0% | Average per-task cost is about $0.20; low tier reaches 72%+ @ $0.10; the previous-generation V4.1-Flash was 74.2% @ $0.50 |
| SWE-Bench Verified | V4-Pro | 80.6% | Fixing real-world repository-level issues |
| LiveCodeBench | V4-Pro | 93.5% | Real-time competitive programming |
| Terminal-Bench 2.0 | V4-Pro | 67.9% | Terminal Agent |
| Codeforces | V4-Pro | 3,206 | Competitive Programming Rating |
| Putnam-200 (Pass@8) | V4-Flash | 81.0% | Math competitions |
| SWE-Bench Verified | V4-Flash | 79.0% | Only 1.6pt behind the flagship, at about 1/4 the price |
Source: DeepSeek official release page and API documentation (V4 Preview 2026-04-24, V4.1-Flash 2026-09-10, 1017 2026-10-17).
Comparison scores officially published at GLM-5.3's release; the DeepSeek-V4-Pro column is highlighted. GPT-5.6 Sol was the strongest in the comparison at the time, while V4-Pro remained in the top tier among open-source models.
| Benchmark | GLM-5.3 | Kimi K3 | DeepSeek-V4-Pro | Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 88.2 | 88.3 | 87.9 | 85.0 | 88.8 |
| DeepSWE v1.1 | 66.9 | 67.5 | 62.7 | 58.0 | 72.7 |
| NL2Repo | 58.0 | 58.0 | 61.1 | 69.7 | — |
| CyberGym | 84.5 | 80.0 | 83.3 | 78.1 | 83.6 |
| Toolathlon Verified | 73.0 | 76.5 | 74.1 | 76.2 | 74.9 |
| AutomationBench v1.0.6 | 48.2 | 46.7 | 43.2 | 41.0 | 45.8 |
| ALE-CLI | 28.5 | 27.6 | 25.7 | 25.7 | 28.6 |
| HLE w/ Tools | 62.5 | 59.8 | 60.0 | 57.9 | 64.5 |
| GDPval-AA v2 | 1769 | 1682 | 1590 | 1588 | 1730 |
Data source: Z.ai official blog (2026-08-14); items not evaluated are indicated with —. Price comparison: V4-Pro off-peak $0.66/$1.98, V4-Flash $0.14/$0.28 per million tokens.
Under 1M token workloads, total compute overhead is approximately 27% of V3.2's — a direct benefit of the hybrid attention (CSA+HCA).
KV cache footprint is approximately 10% of V3.2's, sharply reducing VRAM costs for long-context concurrent serving.
The V4.1-Flash family offers cache-hit input as low as $0.003/1M tokens, bringing pipeline costs close to zero for workflows that repeatedly retrieve the same corpus.
大模型网摘