Model comparison
DeepSeek-V3.2-Exp vs GPT-4
DeepSeek-V3.2-Exp is the stronger model overall, scoring 44.3 to 29.1 on the Noometry Index.
Last verified . 25 shared benchmarks.
DeepSeekSummary
- They share 25 benchmarks with published results for both. DeepSeek-V3.2-Exp scores higher in 8 categories and GPT-4 in 0 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in knowledge, where DeepSeek-V3.2-Exp leads 51.7 to 18.4.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 87.8% for DeepSeek-V3.2-Exp and 1.1% for GPT-4.
- DeepSeek-V3.2-Exp is cheaper at $0.26 / $0.38 per million input/output tokens, against $30 / $60 for GPT-4.
- DeepSeek-V3.2-Exp accepts more context: 164K tokens versus 8K.
- DeepSeek-V3.2-Exp has downloadable open weights; the other is API-only.
Side by side
| DeepSeek-V3.2-Exp | GPT-4 | |
|---|---|---|
| Provider | DeepSeek | OpenAI |
| Noometry Index | 44.3 | 29.1 |
| Released | 2025-09-29 | 2023-03-14 |
| Weights | Open | Proprietary |
| Context window | 164K | 8K |
| Max output | 66K | 8K |
| Input $ / M tokens | $0.26 | $30 |
| Output $ / M tokens | $0.38 | $60 |
| Results tracked | 49 | 38 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek-V3.2-Exp leads
DeepSeek-V3.2-Exp: 46.5 (#65), GPT-4: 31.6 (#283)
| Benchmark | DeepSeek-V3.2-Exp | GPT-4 |
|---|---|---|
| WeirdML | 39.5% | 12.4% |
| LMArena Coding | 1454 | 1254 |
| SWE-bench Verified (bash only) | 70% | — |
| Aider Polyglot | 74.2% | — |
| LMArena WebDev | 1362 | — |
| SWE-bench Multilingual | 59% | — |
| SciCode | 38.9% | — |
| BigCodeBench Instruct | — | 46% |
| BigCodeBench Complete | — | 57.2% |
| HumanEval+ | — | 79.3% |
Agentic & Tool Use Not comparable
DeepSeek-V3.2-Exp: 32.7 (#59), GPT-4: —
| Benchmark | DeepSeek-V3.2-Exp | GPT-4 |
|---|---|---|
| Terminal-Bench | 39.6% | — |
| APEX-Agents | 21.3% | — |
| Berkeley Function Calling Leaderboard | 56.7% | — |
| TheAgentCompany | 42.9% | — |
| METR Time Horizons | — | 36.1% |
| Vending-Bench 2 | 1,034 | — |
Reasoning DeepSeek-V3.2-Exp leads
DeepSeek-V3.2-Exp: 22.1 (#208), GPT-4: 17.8 (#289)
| Benchmark | DeepSeek-V3.2-Exp | GPT-4 |
|---|---|---|
| Chess Puzzles | 14% | 4% |
| LMArena Hard Prompts | 1434 | 1241 |
| DTBench | 87.7% | 62.7% |
| LMCA | 29.1% | 17.1% |
| Epoch Capabilities Index | 146.27 | 125.89 |
| ARC-AGI-2 | 4% | — |
| Kagi LLM Benchmark | 52.2% | — |
| NYT Connections (extended) | 36.7% | — |
| ARC-AGI-1 | 57% | — |
| CritPt | 2.9% | — |
| Thematic Generalization | 65% | — |
| Mystery Game Puzzles | — | 12% |
| BIG-Bench Hard | — | 75.1% |
| ForecastBench | — | 57.8 |
| HellaSwag | — | 95.3% |
| WinoGrande | — | 87.5% |
Math DeepSeek-V3.2-Exp leads
DeepSeek-V3.2-Exp: 41.7 (#87), GPT-4: 10.8 (#309)
| Benchmark | DeepSeek-V3.2-Exp | GPT-4 |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 87.8% | 1.1% |
| LMArena Math | 1435 | 1269 |
| MathArena Final-Answer Competitions | 57.7% | — |
| ProofBench | 8% | — |
| MATH Level 5 | — | 23% |
| FrontierMath (Feb 2025 set) | 22.1% | — |
| FrontierMath Tier 4 (v1) | 2.1% | — |
| GSM8K | — | 92% |
Knowledge DeepSeek-V3.2-Exp leads
DeepSeek-V3.2-Exp: 51.7 (#66), GPT-4: 18.4 (#282)
| Benchmark | DeepSeek-V3.2-Exp | GPT-4 |
|---|---|---|
| GPQA Diamond | 83.4% | 35.7% |
| LMArena Expert | 1436 | 1211 |
| Vectara Hallucination Rate | 5.3% | — |
| MMLU | — | 86.4% |
| TriviaQA | — | 84.8% |
Multilingual DeepSeek-V3.2-Exp leads
DeepSeek-V3.2-Exp: 52.2 (#90), GPT-4: 40.6 (#215)
| Benchmark | DeepSeek-V3.2-Exp | GPT-4 |
|---|---|---|
| LMArena Non-English | 1409 | 1246 |
| LMArena Chinese | 1461 | 1242 |
| LMArena French | 1433 | 1283 |
| LMArena German | 1440 | 1251 |
| LMArena Japanese | 1374 | 1209 |
| LMArena Korean | 1371 | 1184 |
| LMArena Russian | 1424 | 1251 |
| LMArena Spanish | 1440 | 1261 |
Instruction Following DeepSeek-V3.2-Exp leads
DeepSeek-V3.2-Exp: 74.5 (#93), GPT-4: 65.3 (#222)
| Benchmark | DeepSeek-V3.2-Exp | GPT-4 |
|---|---|---|
| LMArena Instruction Following | 1413 | 1241 |
Long Context DeepSeek-V3.2-Exp leads
DeepSeek-V3.2-Exp: 47.6 (#16), GPT-4: 37.7 (#212)
| Benchmark | DeepSeek-V3.2-Exp | GPT-4 |
|---|---|---|
| LMArena Longer Query | 1428 | 1244 |
| Fiction.LiveBench | 83.3% | — |
| CL-bench | 13.2% | — |
| CL-bench Life | 9.5% | — |
Writing & Preference DeepSeek-V3.2-Exp leads
DeepSeek-V3.2-Exp: 62.4 (#77), GPT-4: 34.9 (#268)
| Benchmark | DeepSeek-V3.2-Exp | GPT-4 |
|---|---|---|
| LMArena Text | 1425 | 1263 |
| LMArena Creative Writing | 1403 | 1244 |
| EQ-Bench Creative Writing | 1515 | 752 |
| LMArena Multi-Turn | 1427 | 1257 |
Frequently asked questions
Is DeepSeek-V3.2-Exp better than GPT-4?
DeepSeek-V3.2-Exp is the stronger model overall, scoring 44.3 to 29.1 on the Noometry Index.
Which is cheaper, DeepSeek-V3.2-Exp or GPT-4?
DeepSeek-V3.2-Exp is cheaper. It lists at $0.26 per million input tokens and $0.38 per million output tokens; GPT-4 lists at $30 and $60.
Is DeepSeek-V3.2-Exp or GPT-4 better for coding?
DeepSeek-V3.2-Exp scores higher on coding benchmarks: 46.5 versus 31.6 in the Noometry coding category.
Which has the bigger context window?
DeepSeek-V3.2-Exp does, with 164K tokens against 8K.
How many benchmarks do DeepSeek-V3.2-Exp and GPT-4 share?
25 benchmarks have published results for both models. DeepSeek-V3.2-Exp has 49 scored results on Noometry and GPT-4 has 38.