- Published
- Sep 16, 2026, 06:28 AM UTC
- Question pack
- coding-fast-v4.12
- Tested
- 4 measured configurations
- Batch
- evaluation-497e806be8b8470345169c1aad6752d74a0c1ce9e195a9eb7fcc02b628701c77
Measured configurations
| Configuration | Rank | Score | Elapsed | Reference cost | Route |
|---|---|---|---|---|---|
| DeepSeek V4 Pro / Max | #30 | 59/100 | 30m 2s | $0.41 | Custom endpoint |
| DeepSeek V4 Flash / High | #36 | 54/100 | 12m 26s | $0.0830 | Custom endpoint |
| DeepSeek V4 Pro / High | #43 | 50/100 | 27m 3s | $0.33 | Custom endpoint |
| DeepSeek V4 Flash / Max | #45 | 47/100 | 13m 46s | $0.10 | Custom endpoint |
Five-question profile
| Question | DeepSeek V4 Pro / Max | DeepSeek V4 Flash / High | DeepSeek V4 Pro / High | DeepSeek V4 Flash / Max |
|---|---|---|---|---|
| Black-box Regression Audit | 10/20 | 8/20 | 5/20 | 11/20 |
| Retry Planner Counterexamples | 15/20 | 15/20 | 16/20 | 13/20 |
| CI Adversarial Audit | 11/20 | 10/20 | 7/20 | 14/20 |
| Transaction Regression Design | 12/20 | 9/20 | 13/20 | 0/20 |
| Cache Propagation Certificate | 11/20 | 12/20 | 9/20 | 9/20 |
Flash and Pro are not interchangeable rows
DeepSeek V4 Flash and Pro are distinct model identities. Each effort level is another configuration. Combining them into one DeepSeek score would hide which variant produced the result.
Use the table to find whether quality, elapsed time, and cost point to the same configuration. When they do not, apply a task-specific quality floor before optimizing the other two.