# DeepSeek-V3 vs DeepSeek V4 Pro

> DeepSeek V4 Pro is the stronger model overall, scoring 54.2 to 39.2 on the Noometry Index.

- Canonical page: https://noometry.com/compare/deepseek-v3-vs-deepseek-v4-pro
- Last updated: 2026-10-10
- Shared benchmarks: 29

## Summary

- They share 29 benchmarks with published results for both. DeepSeek-V3 scores higher in 0 categories and DeepSeek V4 Pro in 8 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in reasoning, where DeepSeek V4 Pro leads 56.4 to 20.0.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 37.8% for DeepSeek-V3 and 98.6% for DeepSeek V4 Pro.
- DeepSeek-V3 is cheaper at $0.24 / $0.90 per million input/output tokens, against $0.43 / $0.87 for DeepSeek V4 Pro.
- DeepSeek V4 Pro accepts more context: 1M tokens versus 164K.

## Snapshot

| | DeepSeek-V3 | DeepSeek V4 Pro |
|---|---|---|
| Provider | DeepSeek | DeepSeek |
| Noometry Index | 39.2 | 54.2 |
| Rank | 157 | 28 |
| Context | 164K | 1M |
| Input $/M | $0.24 | $0.43 |
| Output $/M | $0.90 | $0.87 |
| Weights | Open | Open |

## Coding

- DeepSeek-V3: 41.9 (#99)
- DeepSeek V4 Pro: 52.2 (#31)

| Benchmark | DeepSeek-V3 | DeepSeek V4 Pro |
|---|---|---|
| SciCode | 35.8% | 51% |
| WeirdML | 36.1% | 66.2% |
| LMArena Coding | 1368 | 1470 |
| SWE-bench Verified | — | 77.6% |
| FrontierCode | — | 28.6% |
| Aider Polyglot | 55.1% | — |
| LMArena WebDev | — | 1582 |
| BigCodeBench Instruct | 50% | — |
| LiveBench Coding | 70.9% | — |
| BigCodeBench Complete | 62.2% | — |
| ALE-Bench | — | 1,403 |
| HumanEval+ | 86.6% | — |
| MBPP+ | 73% | — |

## Agentic & Tool Use

- DeepSeek-V3: —
- DeepSeek V4 Pro: 32.7 (#54)

| Benchmark | DeepSeek-V3 | DeepSeek V4 Pro |
|---|---|---|
| APEX-Agents | — | 47.3% |
| METR Time Horizons | 49.6% | — |
| Vending-Bench 2 | — | 3,285 |

## Reasoning

- DeepSeek-V3: 20.0 (#226)
- DeepSeek V4 Pro: 56.4 (#22)

| Benchmark | DeepSeek-V3 | DeepSeek V4 Pro |
|---|---|---|
| Kagi LLM Benchmark | 52.3% | 53.5% |
| CritPt | 0% | 18% |
| LMArena Hard Prompts | 1365 | 1461 |
| DTBench | 64.8% | 93.9% |
| LMCA | 15.5% | 45.5% |
| Epoch Capabilities Index | 135.95 | 155.39 |
| ForecastBench | 59.1 | 56.1 |
| ARC-AGI-2 | — | 61.3% |
| SimpleBench | 27.2% | — |
| NYT Connections (extended) | — | 91.3% |
| ARC-AGI-1 | — | 90.5% |
| Chess Puzzles | — | 47% |
| LiveBench Reasoning | 65.8% | — |
| Mystery Game Puzzles | — | 43% |
| LiveBench Data Analysis | 60.9% | — |
| Surface Evolver Bench | — | 40% |
| BIG-Bench Hard | 87.5% | — |
| HellaSwag | 88.9% | — |
| LiveBench | 66.9% | — |
| PIQA | 84.7% | — |
| WinoGrande | 85.2% | — |

## Math

- DeepSeek-V3: 30.4 (#208)
- DeepSeek V4 Pro: 64.6 (#24)

| Benchmark | DeepSeek-V3 | DeepSeek V4 Pro |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 37.8% | 98.6% |
| LMArena Math | 1373 | 1455 |
| FrontierMath (Tiers 1-3) | — | 64.6% |
| FrontierMath Tier 4 | — | 26.8% |
| MathArena Final-Answer Competitions | — | 76.6% |
| ProofBench | — | 50% |
| Omni-MATH | 40.3% | — |
| LiveBench Math | 73.5% | — |
| MATH Level 5 | 75.5% | — |
| FrontierMath (Feb 2025 set) | 1.7% | — |

## Knowledge

- DeepSeek-V3: 37.5 (#147)
- DeepSeek V4 Pro: 59.5 (#28)

| Benchmark | DeepSeek-V3 | DeepSeek V4 Pro |
|---|---|---|
| GPQA Diamond | 67.6% | 91.7% |
| Vectara Hallucination Rate | 6.1% | 8.6% |
| LMArena Expert | 1352 | 1463 |
| SimpleQA Verified | — | 52.9% |
| MMLU-Pro | 72.3% | — |
| Confabulations | 26.1% | — |
| GPQA (HELM) | 53.8% | — |
| ARC (AI2) Challenge | 95.3% | — |
| MMLU | 87.2% | — |
| TriviaQA | 82.9% | — |

## Multilingual

- DeepSeek-V3: 48.8 (#138)
- DeepSeek V4 Pro: 54.5 (#44)

| Benchmark | DeepSeek-V3 | DeepSeek V4 Pro |
|---|---|---|
| LMArena Non-English | 1358 | 1438 |
| LMArena Chinese | 1390 | 1486 |
| LMArena French | 1386 | 1473 |
| LMArena German | 1374 | 1458 |
| LMArena Japanese | 1333 | 1448 |
| LMArena Korean | 1321 | 1452 |
| LMArena Russian | 1372 | 1452 |
| LMArena Spanish | 1358 | 1457 |

## Instruction Following

- DeepSeek-V3: 73.0 (#125)
- DeepSeek V4 Pro: 76.3 (#44)

| Benchmark | DeepSeek-V3 | DeepSeek V4 Pro |
|---|---|---|
| LMArena Instruction Following | 1345 | 1448 |
| LiveBench Instruction Following | 81.5% | — |
| IFEval | 83.2% | — |

## Long Context

- DeepSeek-V3: 34.0 (#248)
- DeepSeek V4 Pro: 45.0 (#49)

| Benchmark | DeepSeek-V3 | DeepSeek V4 Pro |
|---|---|---|
| LMArena Longer Query | 1353 | 1459 |
| Fiction.LiveBench | 50% | — |
| CL-bench Life | — | 13.5% |

## Writing & Preference

- DeepSeek-V3: 57.6 (#124)
- DeepSeek V4 Pro: 65.6 (#45)

| Benchmark | DeepSeek-V3 | DeepSeek V4 Pro |
|---|---|---|
| LMArena Text | 1375 | 1451 |
| LMArena Creative Writing | 1364 | 1446 |
| EQ-Bench Creative Writing | 1472 | 1553 |
| LMArena Multi-Turn | 1389 | 1467 |
| Short-Story Creative Writing | 77% | — |
| WildBench | 83% | — |
| EQ-Bench 4 | — | 1166 |
| LiveBench Language | 49.1% | — |

## FAQ

### Is DeepSeek-V3 better than DeepSeek V4 Pro?

DeepSeek V4 Pro is the stronger model overall, scoring 54.2 to 39.2 on the Noometry Index.

### Which is cheaper, DeepSeek-V3 or DeepSeek V4 Pro?

DeepSeek-V3 is cheaper. It lists at $0.24 per million input tokens and $0.90 per million output tokens; DeepSeek V4 Pro lists at $0.43 and $0.87.

### Is DeepSeek-V3 or DeepSeek V4 Pro better for coding?

DeepSeek V4 Pro scores higher on coding benchmarks: 52.2 versus 41.9 in the Noometry coding category.

### Which has the bigger context window?

DeepSeek V4 Pro does, with 1M tokens against 164K.

### How many benchmarks do DeepSeek-V3 and DeepSeek V4 Pro share?

29 benchmarks have published results for both models. DeepSeek-V3 has 60 scored results on Noometry and DeepSeek V4 Pro has 48.