In short
How they differ.
Worked out from the figures below. Prices and specs are the makers' own, read Oct 20, 2026.
- DeepSeek-V4.1-Flash-1017 is cheaper per output token: $1.20 against $3.96 per million tokens, so DeepSeek-V4-Pro-0813 costs 3.3 times as much.
- Both can be downloaded and run on your own machine.
- On the 1 independent board that rank both, DeepSeek-V4.1-Flash-1017 places higher on 1 and DeepSeek-V4-Pro-0813 on 0.
Side by side
The specs.
| DeepSeek-V4.1-Flash-1017 | DeepSeek-V4-Pro-0813 | |
|---|---|---|
| Made by | DeepSeek · China | DeepSeek · China |
| Released | Oct 17, 2026 | Aug 13, 2026 |
| Price per 1M tokens | $0.30 in / $1.20 out (peak-hours price) | $1.32 in / $3.96 out (peak-hours price) |
| Context window | 1M | 1M |
| Takes in | text, image | text |
| Download and run it yourself | Yes (MIT) | Yes (MIT) |
| Size | 552B backbone parameters parameters, 8B per token during prefill, 16B during decode active | 1.6T parameters, 49B active |
| Maker says | “Stronger benchmark performance, faster task completion, and shorter outputs than the previous version.” | — |
| Sources | Prices, Oct 20, 2026 · Description | Prices, Oct 20, 2026 |
Independent tests
Where they place.
Position on each board we follow (1 is best). Epoch AI's and Vectara's scores are shown; the other boards are linked for their own figures.
| Board | DeepSeek-V4.1-Flash-1017 | DeepSeek-V4-Pro-0813 |
|---|---|---|
| Humanity's Last ExamScale AI | Not listed | #16 |
| Vals IndexVals AI, Oct 19, 2026 | #8 | #17 |
| Terminal-Bench 4.0Vals AI, Oct 19, 2026 | #17 | Not listed |
| Vibe Code BenchVals AI, Oct 18, 2026 | #11 | Not listed |
| GPQA Diamond (run by Vals AI, archived)Vals AI, Sep 1, 2026 | Not listed | #17 |
| Vectara hallucination testVectara, Oct 16, 2026 | Not listed | #448.6% |
More comparisons