DeepSeek-V4-ProvsDeepSeek-V3.1
Across 5 shared benchmarks, DeepSeek-V3.1 leads overall: DeepSeek-V4-Pro wins 1, DeepSeek-V3.1 wins 4, with 0 ties and an average score difference of -5.58.
DeepSeek-V4-Pro
DeepSeek-AI · 2026-04-24 · Reasoning model
DeepSeek-V3.1
DeepSeek-AI · 2025-08-20 · Chat model
DeepSeek-V4-Pro1 win(20%)(80%)4 winsDeepSeek-V3.1
Benchmark scores
Grouped by capability, sorted by largest gap within each. 5 shared benchmarks.
General Knowledge
DeepSeek-V3.1 3/3| Benchmark | DeepSeek-V4-Pro | DeepSeek-V3.1 | Diff |
|---|---|---|---|
| HLE | 7.70141 / 157Normal (No Tools) | 15.90118 / 157 | -8.20 |
| GPQA Diamond | 72.90102 / 178Normal (No Tools) | 80.1075 / 178 | -7.20 |
| MMLU Pro | 82.9046 / 126Normal (No Tools) | 8525 / 126 | -2.10 |
Coding and Software Engineer
Even 2/2| Benchmark | DeepSeek-V4-Pro | DeepSeek-V3.1 | Diff |
|---|---|---|---|
| LiveCodeBench | 56.8075 / 120Normal (No Tools) | 74.8040 / 120 | -18 |
| SWE-bench Verified | 73.6041 / 108Normal (With Tools) | 6670 / 108 | +7.60 |
Specs
| Field | DeepSeek-V4-Pro | DeepSeek-V3.1 |
|---|---|---|
| Publisher | DeepSeek-AI | DeepSeek-AI |
| Release date | 2026-04-24 | 2025-08-20 |
| Model type | Reasoning model | Chat model |
| Architecture | MoE | MoE |
| Parameters | 1.6T | 671B |
| Context length | 1M | 128K |
| Max output | 375K | 8K |
API pricing
Prices use DataLearner records when available; missing fields are not inferred.
| Item | DeepSeek-V4-Pro | DeepSeek-V3.1 |
|---|---|---|
| Text input | $0.435 / 1M tokens | Not public |
| Text output | $0.87 / 1M tokens | Not public |
| Cache read | $0.87 / 1M tokens | Not public |
| Cache write | $0.003625 / 1M tokens | Not public |
One or both models have incomplete public pricing.
Summary
- DeepSeek-V3.1leads in:General Knowledge (3/3)
- Tied in:Coding and Software Engineer
On average across the 5 shared benchmarks, DeepSeek-V3.1 scores 5.58 higher.
Largest single-benchmark gap: LiveCodeBench — DeepSeek-V4-Pro 56.80 vs DeepSeek-V3.1 74.80 (-18).
Page generated from structured model, pricing and benchmark records. No real-time LLM is used to write the prose.