# DeepSeek-V3.2-Speciale vs GPT-4

> DeepSeek-V3.2-Speciale is the stronger model overall, scoring 39.1 to 28.9 on the Noometry Index.

- Canonical page: https://noometry.com/compare/deepseek-v3-2-speciale-vs-gpt-4
- Last updated: 2026-10-10
- Shared benchmarks: 2

## Summary

- They share 2 benchmarks with published results for both. DeepSeek-V3.2-Speciale scores higher in 3 categories and GPT-4 in 0 categories; 3 gaps are clear of the uncertainty.
- The widest gap is in reasoning, where DeepSeek-V3.2-Speciale leads 32.1 to 17.3.
- The biggest single-benchmark swing is WeirdML: 46.7% for DeepSeek-V3.2-Speciale and 12.4% for GPT-4.
- DeepSeek-V3.2-Speciale is cheaper at $0.58 / $1.68 per million input/output tokens, against $30 / $60 for GPT-4.
- DeepSeek-V3.2-Speciale accepts more context: 128K tokens versus 8K.
- DeepSeek-V3.2-Speciale has downloadable open weights; the other is API-only.

## Snapshot

| | DeepSeek-V3.2-Speciale | GPT-4 |
|---|---|---|
| Provider | DeepSeek | OpenAI |
| Noometry Index | 39.1 | 28.9 |
| Rank | 159 | 308 |
| Context | 128K | 8K |
| Input $/M | $0.58 | $30 |
| Output $/M | $1.68 | $60 |
| Weights | Open | Proprietary |

## Coding

- DeepSeek-V3.2-Speciale: 39.9 (#136)
- GPT-4: 31.2 (#277)

| Benchmark | DeepSeek-V3.2-Speciale | GPT-4 |
|---|---|---|
| WeirdML | 46.7% | 12.4% |
| BigCodeBench Instruct | — | 46% |
| LMArena Coding | — | 1254 |
| BigCodeBench Complete | — | 57.2% |
| HumanEval+ | — | 79.3% |

## Agentic & Tool Use

- DeepSeek-V3.2-Speciale: —
- GPT-4: —

| Benchmark | DeepSeek-V3.2-Speciale | GPT-4 |
|---|---|---|
| METR Time Horizons | — | 36.1% |

## Reasoning

- DeepSeek-V3.2-Speciale: 32.1 (#65)
- GPT-4: 17.3 (#285)

| Benchmark | DeepSeek-V3.2-Speciale | GPT-4 |
|---|---|---|
| SimpleBench | 52.6% | — |
| Chess Puzzles | — | 4% |
| LMArena Hard Prompts | — | 1241 |
| Mystery Game Puzzles | — | 12% |
| DTBench | — | 62.7% |
| LMCA | — | 17.1% |
| BIG-Bench Hard | — | 75.1% |
| Epoch Capabilities Index | — | 125.88 |
| ForecastBench | — | 57.8 |
| HellaSwag | — | 95.3% |
| WinoGrande | — | 87.5% |

## Math

- DeepSeek-V3.2-Speciale: —
- GPT-4: 10.1 (#304)

| Benchmark | DeepSeek-V3.2-Speciale | GPT-4 |
|---|---|---|
| OTIS Mock AIME 2024-2025 | — | 1.1% |
| LMArena Math | — | 1269 |
| MATH Level 5 | — | 23% |
| GSM8K | — | 92% |

## Knowledge

- DeepSeek-V3.2-Speciale: —
- GPT-4: 18.4 (#277)

| Benchmark | DeepSeek-V3.2-Speciale | GPT-4 |
|---|---|---|
| GPQA Diamond | — | 35.7% |
| LMArena Expert | — | 1211 |
| MMLU | — | 86.4% |
| TriviaQA | — | 84.8% |

## Multilingual

- DeepSeek-V3.2-Speciale: —
- GPT-4: 40.8 (#210)

| Benchmark | DeepSeek-V3.2-Speciale | GPT-4 |
|---|---|---|
| LMArena Non-English | — | 1246 |
| LMArena Chinese | — | 1242 |
| LMArena French | — | 1282 |
| LMArena German | — | 1251 |
| LMArena Japanese | — | 1209 |
| LMArena Korean | — | 1184 |
| LMArena Russian | — | 1251 |
| LMArena Spanish | — | 1261 |

## Instruction Following

- DeepSeek-V3.2-Speciale: —
- GPT-4: 65.5 (#217)

| Benchmark | DeepSeek-V3.2-Speciale | GPT-4 |
|---|---|---|
| LMArena Instruction Following | — | 1241 |

## Long Context

- DeepSeek-V3.2-Speciale: —
- GPT-4: 37.7 (#207)

| Benchmark | DeepSeek-V3.2-Speciale | GPT-4 |
|---|---|---|
| LMArena Longer Query | — | 1244 |

## Writing & Preference

- DeepSeek-V3.2-Speciale: 46.1 (#217)
- GPT-4: 35.0 (#263)

| Benchmark | DeepSeek-V3.2-Speciale | GPT-4 |
|---|---|---|
| EQ-Bench Creative Writing | 1276 | 752 |
| LMArena Text | — | 1263 |
| LMArena Creative Writing | — | 1244 |
| LMArena Multi-Turn | — | 1257 |

## FAQ

### Is DeepSeek-V3.2-Speciale better than GPT-4?

DeepSeek-V3.2-Speciale is the stronger model overall, scoring 39.1 to 28.9 on the Noometry Index.

### Which is cheaper, DeepSeek-V3.2-Speciale or GPT-4?

DeepSeek-V3.2-Speciale is cheaper. It lists at $0.58 per million input tokens and $1.68 per million output tokens; GPT-4 lists at $30 and $60.

### Is DeepSeek-V3.2-Speciale or GPT-4 better for coding?

DeepSeek-V3.2-Speciale scores higher on coding benchmarks: 39.9 versus 31.2 in the Noometry coding category.

### Which has the bigger context window?

DeepSeek-V3.2-Speciale does, with 128K tokens against 8K.

### How many benchmarks do DeepSeek-V3.2-Speciale and GPT-4 share?

2 benchmarks have published results for both models. DeepSeek-V3.2-Speciale has 3 scored results on Noometry and GPT-4 has 38.