Head to head
deepseek-flash vs deepseek-v4-pro
Across 8 tests, deepseek-flash and deepseek-v4-pro are not clearly apart on 6. deepseek-flash is ahead on 1 and deepseek-v4-pro on 1. deepseek-flash costs less on 8 of the 8.
A model is shown ahead only when the data says it is probably better. Otherwise the result is "not clearly apart".
Right answers, test by test
Each row is a task and difficulty, with each model at its best thinking setting. The same tests at list price, log scale.
The numbers behind this chart are in the table below it.
| Test | deepseek-flash: Thinking | Right answers | Per 1,000 tasks | deepseek-v4-pro: Thinking | Right answers | Per 1,000 tasks | Result |
|---|---|---|---|---|---|---|---|
| Math, Easy | Thinking low | 100.0% | $0.27 | Thinking low | 100.0% | $0.81 | Not clearly apart |
| Math, Medium | Thinking low | 98.2% | $0.62 | Thinking low | 98.2% | $1.57 | Not clearly apart |
| Math, Hard | Thinking low | 100.0% | $2.26 | Thinking high | 98.2% | $15.63 | Not clearly apart |
| Math, Hardest | Thinking low | 96.4% | $6.38 | Thinking high | 98.2% | $48.68 | deepseek-v4-pro ahead |
| SQL and data queries, Easy | Thinking off | 100.0% | $0.24 | Thinking off | 100.0% | $1.03 | Not clearly apart |
| SQL and data queries, Medium | Thinking high | 98.2% | $1.87 | Thinking low | 98.2% | $8.92 | Not clearly apart |
| SQL and data queries, Hard | Thinking low | 100.0% | $2.10 | Thinking high | 100.0% | $22.80 | Not clearly apart |
| SQL and data queries, Hardest | Thinking low | 100.0% | $4.29 | Thinking high | 98.2% | $41.95 | deepseek-flash ahead |
Read this first
- AIME problems are public and may appear in model training data, so math results at the frontier level may be inflated. Differences between settings are the reliable part.
- SQL questions were kept only when an independent model outside the test answered them correctly, which limits how hard the hardest questions are.
- These results compare DeepSeek models and settings with each other. Other providers' models have not been tested in this run.
- Results are not ranked when their ranges overlap. Top means nothing tested is probably better on that job. Near the top means not clearly worse than the top.
Zabio sells access to these models and earns a fee on the traffic it carries. That is why the method is public, every number has its range and date, and corrections are published as changes. No vendor pays for tests, results or placement.
DeepSeek is a trademark of its owner. Zabio uses DeepSeek models through the DeepSeek API and is not affiliated with DeepSeek.
Use the data
The results are free to reuse under CC BY 4.0. Please credit Zabio with the citation below.
- License
- CC BY 4.0
- Suggested citation
Zabio, Model benchmarks, zabio.ai/benchmarks, 2026-09-25- Methodology
- 1.0
- Tested
- 25 September 2026