Model comparison
DeepSeek-V3 vs GPT-4
DeepSeek-V3 is the stronger model overall, scoring 39.5 to 29.1 on the Noometry Index.
Last verified . 35 shared benchmarks.
DeepSeekSummary
- They share 35 benchmarks with published results for both. DeepSeek-V3 scores higher in 7 categories and GPT-4 in 1 category; 8 gaps are clear of the uncertainty.
- The widest gap is in writing & preference, where DeepSeek-V3 leads 57.4 to 34.9.
- The biggest single-benchmark swing is MATH Level 5: 75.5% for DeepSeek-V3 and 23% for GPT-4.
- DeepSeek-V3 is cheaper at $0.24 / $0.90 per million input/output tokens, against $30 / $60 for GPT-4.
- DeepSeek-V3 accepts more context: 164K tokens versus 8K.
- DeepSeek-V3 has downloadable open weights; the other is API-only.
Side by side
| DeepSeek-V3 | GPT-4 | |
|---|---|---|
| Provider | DeepSeek | OpenAI |
| Noometry Index | 39.5 | 29.1 |
| Released | 2024-12-26 | 2023-03-14 |
| Weights | Open | Proprietary |
| Context window | 164K | 8K |
| Max output | 164K | 8K |
| Input $ / M tokens | $0.24 | $30 |
| Output $ / M tokens | $0.90 | $60 |
| Results tracked | 60 | 38 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek-V3 leads
DeepSeek-V3: 42.3 (#106), GPT-4: 31.6 (#283)
| Benchmark | DeepSeek-V3 | GPT-4 |
|---|---|---|
| WeirdML | 36.1% | 12.4% |
| BigCodeBench Instruct | 50% | 46% |
| LMArena Coding | 1368 | 1254 |
| BigCodeBench Complete | 62.2% | 57.2% |
| HumanEval+ | 86.6% | 79.3% |
| Aider Polyglot | 55.1% | — |
| SciCode | 35.8% | — |
| LiveBench Coding | 70.9% | — |
| MBPP+ | 73% | — |
Agentic & Tool Use Not comparable
DeepSeek-V3: —, GPT-4: —
| Benchmark | DeepSeek-V3 | GPT-4 |
|---|---|---|
| METR Time Horizons | 49.6% | 36.1% |
Reasoning DeepSeek-V3 leads
DeepSeek-V3: 20.5 (#236), GPT-4: 17.8 (#289)
| Benchmark | DeepSeek-V3 | GPT-4 |
|---|---|---|
| LMArena Hard Prompts | 1365 | 1241 |
| DTBench | 64.8% | 62.7% |
| LMCA | 15.5% | 17.1% |
| BIG-Bench Hard | 87.5% | 75.1% |
| Epoch Capabilities Index | 135.94 | 125.89 |
| ForecastBench | 59.1 | 57.8 |
| HellaSwag | 88.9% | 95.3% |
| WinoGrande | 85.2% | 87.5% |
| SimpleBench | 27.2% | — |
| Kagi LLM Benchmark | 52.3% | — |
| CritPt | 0% | — |
| Chess Puzzles | — | 4% |
| LiveBench Reasoning | 65.8% | — |
| Mystery Game Puzzles | — | 12% |
| LiveBench Data Analysis | 60.9% | — |
| LiveBench | 66.9% | — |
| PIQA | 84.7% | — |
Math DeepSeek-V3 leads
DeepSeek-V3: 32.1 (#219), GPT-4: 10.8 (#309)
| Benchmark | DeepSeek-V3 | GPT-4 |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 37.8% | 1.1% |
| LMArena Math | 1373 | 1269 |
| MATH Level 5 | 75.5% | 23% |
| Omni-MATH | 40.3% | — |
| LiveBench Math | 73.5% | — |
| FrontierMath (Feb 2025 set) | 1.7% | — |
| GSM8K | — | 92% |
Knowledge DeepSeek-V3 leads
DeepSeek-V3: 37.5 (#155), GPT-4: 18.4 (#282)
| Benchmark | DeepSeek-V3 | GPT-4 |
|---|---|---|
| GPQA Diamond | 67.6% | 35.7% |
| LMArena Expert | 1351 | 1211 |
| MMLU | 87.2% | 86.4% |
| TriviaQA | 82.9% | 84.8% |
| MMLU-Pro | 72.3% | — |
| Confabulations | 26.1% | — |
| Vectara Hallucination Rate | 6.1% | — |
| GPQA (HELM) | 53.8% | — |
| ARC (AI2) Challenge | 95.3% | — |
Multilingual DeepSeek-V3 leads
DeepSeek-V3: 48.5 (#143), GPT-4: 40.6 (#215)
| Benchmark | DeepSeek-V3 | GPT-4 |
|---|---|---|
| LMArena Non-English | 1358 | 1246 |
| LMArena Chinese | 1391 | 1242 |
| LMArena French | 1385 | 1283 |
| LMArena German | 1374 | 1251 |
| LMArena Japanese | 1333 | 1209 |
| LMArena Korean | 1319 | 1184 |
| LMArena Russian | 1373 | 1251 |
| LMArena Spanish | 1358 | 1261 |
Instruction Following DeepSeek-V3 leads
DeepSeek-V3: 72.8 (#130), GPT-4: 65.3 (#222)
| Benchmark | DeepSeek-V3 | GPT-4 |
|---|---|---|
| LMArena Instruction Following | 1345 | 1241 |
| LiveBench Instruction Following | 81.5% | — |
| IFEval | 83.2% | — |
Long Context GPT-4 leads
DeepSeek-V3: 34.0 (#253), GPT-4: 37.7 (#212)
| Benchmark | DeepSeek-V3 | GPT-4 |
|---|---|---|
| LMArena Longer Query | 1352 | 1244 |
| Fiction.LiveBench | 50% | — |
Writing & Preference DeepSeek-V3 leads
DeepSeek-V3: 57.4 (#130), GPT-4: 34.9 (#268)
| Benchmark | DeepSeek-V3 | GPT-4 |
|---|---|---|
| LMArena Text | 1375 | 1263 |
| LMArena Creative Writing | 1364 | 1244 |
| EQ-Bench Creative Writing | 1472 | 752 |
| LMArena Multi-Turn | 1389 | 1257 |
| Short-Story Creative Writing | 77% | — |
| WildBench | 83% | — |
| LiveBench Language | 49.1% | — |
Frequently asked questions
Is DeepSeek-V3 better than GPT-4?
DeepSeek-V3 is the stronger model overall, scoring 39.5 to 29.1 on the Noometry Index.
Which is cheaper, DeepSeek-V3 or GPT-4?
DeepSeek-V3 is cheaper. It lists at $0.24 per million input tokens and $0.90 per million output tokens; GPT-4 lists at $30 and $60.
Is DeepSeek-V3 or GPT-4 better for coding?
DeepSeek-V3 scores higher on coding benchmarks: 42.3 versus 31.6 in the Noometry coding category.
Which has the bigger context window?
DeepSeek-V3 does, with 164K tokens against 8K.
How many benchmarks do DeepSeek-V3 and GPT-4 share?
35 benchmarks have published results for both models. DeepSeek-V3 has 60 scored results on Noometry and GPT-4 has 38.