Model comparison
DeepSeek-V3.1 vs GPT-5.2
GPT-5.2 is the stronger model overall, scoring 54.1 to 42.8 on the Noometry Index. DeepSeek-V3.1 costs 11× less per token, which makes it the better buy when GPT-5.2's lead doesn't matter for your workload.
Last verified . 26 shared benchmarks.
DeepSeekSummary
- They share 26 benchmarks with published results for both. DeepSeek-V3.1 scores higher in 0 categories and GPT-5.2 in 8 categories; 7 gaps are clear of the uncertainty.
- The widest gap is in reasoning, where GPT-5.2 leads 50.2 to 27.9.
- The biggest single-benchmark swing is WeirdML: 38.4% for DeepSeek-V3.1 and 72.2% for GPT-5.2.
- DeepSeek-V3.1 is cheaper at $0.25 / $0.95 per million input/output tokens, against $1.75 / $14 for GPT-5.2.
- GPT-5.2 accepts more context: 400K tokens versus 164K.
- DeepSeek-V3.1 has downloadable open weights; the other is API-only.
Side by side
| DeepSeek-V3.1 | GPT-5.2 | |
|---|---|---|
| Provider | DeepSeek | OpenAI |
| Noometry Index | 42.8 | 54.1 |
| Released | 2025-08-21 | 2025-12-11 |
| Weights | Open | Proprietary |
| Context window | 164K | 400K |
| Max output | 8K | 128K |
| Input $ / M tokens | $0.25 | $1.75 |
| Output $ / M tokens | $0.95 | $14 |
| Results tracked | 27 | 67 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding GPT-5.2 leads
DeepSeek-V3.1: 40.3 (#144), GPT-5.2: 51.6 (#37)
| Benchmark | DeepSeek-V3.1 | GPT-5.2 |
|---|---|---|
| WeirdML | 38.4% | 72.2% |
| LMArena Coding | 1417 | 1447 |
| SWE-bench Verified | — | 73.8% |
| SWE-bench Verified (bash only) | — | 72.8% |
| LMArena WebDev | — | 1416 |
| SWE-bench Multilingual | — | 66.7% |
| GSO | — | 27.4% |
| ALE-Bench | — | 1,294 |
| AlgoTune | — | 2.05 |
Agentic & Tool Use Not comparable
DeepSeek-V3.1: —, GPT-5.2: 40.2 (#24)
| Benchmark | DeepSeek-V3.1 | GPT-5.2 |
|---|---|---|
| Terminal-Bench | — | 64.9% |
| Berkeley Function Calling Leaderboard | — | 55.9% |
| GDPval | — | 49.7% |
| Remote Labor Index | — | 2.5% |
| τ²-bench Airline | — | 83% |
| τ²-bench Banking | — | 32.2% |
| τ²-bench Retail | — | 81.6% |
| τ²-bench Telecom | — | 89.7% |
| DeepResearch Bench | — | 41.1% |
| LMArena Search | — | 1207 |
| METR Time Horizons | — | 75.3% |
| Vending-Bench 2 | — | 3,591 |
Reasoning GPT-5.2 leads
DeepSeek-V3.1: 27.9 (#110), GPT-5.2: 50.2 (#35)
| Benchmark | DeepSeek-V3.1 | GPT-5.2 |
|---|---|---|
| SimpleBench | 40% | 45.8% |
| Kagi LLM Benchmark | 53.2% | 73.3% |
| LMArena Hard Prompts | 1417 | 1445 |
| DTBench | 82.7% | 90.9% |
| LMCA | 24.3% | 43.9% |
| Epoch Capabilities Index | 139.92 | 153.45 |
| ForecastBench | 58 | 60.1 |
| ARC-AGI-2 | — | 52.9% |
| NYT Connections (extended) | — | 83.6% |
| ARC-AGI-1 | — | 86.2% |
| Chess Puzzles | — | 49% |
| EnigmaEval | — | 10.4% |
| EBR-Bench | — | 23% |
| Mystery Game Puzzles | — | 23% |
Math GPT-5.2 leads
DeepSeek-V3.1: 38.9 (#122), GPT-5.2: 60.0 (#38)
| Benchmark | DeepSeek-V3.1 | GPT-5.2 |
|---|---|---|
| LMArena Math | 1420 | 1440 |
| FrontierMath (Tiers 1-3) | — | 67.4% |
| FrontierMath Tier 4 | — | 31.7% |
| MathArena Final-Answer Competitions | — | 72% |
| OTIS Mock AIME 2024-2025 | — | 96.1% |
| ProofBench | — | 15% |
| FrontierMath (Feb 2025 set) | — | 40.7% |
| FrontierMath Tier 4 (v1) | — | 18.8% |
Knowledge GPT-5.2 leads
DeepSeek-V3.1: 43.7 (#90), GPT-5.2: 59.3 (#32)
| Benchmark | DeepSeek-V3.1 | GPT-5.2 |
|---|---|---|
| Vectara Hallucination Rate | 5.5% | 8.4% |
| LMArena Expert | 1405 | 1445 |
| GPQA Diamond | — | 91.4% |
| Humanity's Last Exam | — | 27.8% |
| SimpleQA Verified | — | 37.1% |
Multimodal Not comparable
DeepSeek-V3.1: —, GPT-5.2: 51.3 (#7)
| Benchmark | DeepSeek-V3.1 | GPT-5.2 |
|---|---|---|
| LMArena Vision | — | 1268 |
| VPCT | — | 84% |
| Furniture Assembly | — | 38.3% |
| LMArena Document | — | 1405 |
Multilingual GPT-5.2 leads
DeepSeek-V3.1: 51.6 (#106), GPT-5.2: 53.4 (#67)
| Benchmark | DeepSeek-V3.1 | GPT-5.2 |
|---|---|---|
| LMArena Non-English | 1400 | 1425 |
| LMArena Chinese | 1469 | 1460 |
| LMArena French | 1447 | 1455 |
| LMArena German | 1411 | 1448 |
| LMArena Japanese | 1378 | 1420 |
| LMArena Korean | 1337 | 1392 |
| LMArena Russian | 1405 | 1440 |
| LMArena Spanish | 1431 | 1433 |
Instruction Following Too close to call
DeepSeek-V3.1: 73.9 (#110), GPT-5.2: 74.7 (#89)
| Benchmark | DeepSeek-V3.1 | GPT-5.2 |
|---|---|---|
| LMArena Instruction Following | 1400 | 1417 |
Long Context GPT-5.2 leads
DeepSeek-V3.1: 36.3 (#232), GPT-5.2: 44.0 (#78)
| Benchmark | DeepSeek-V3.1 | GPT-5.2 |
|---|---|---|
| LMArena Longer Query | 1422 | 1428 |
| Fiction.LiveBench | 52.8% | — |
| CL-bench | — | 18.2% |
Writing & Preference GPT-5.2 leads
DeepSeek-V3.1: 60.3 (#98), GPT-5.2: 66.8 (#32)
| Benchmark | DeepSeek-V3.1 | GPT-5.2 |
|---|---|---|
| LMArena Text | 1420 | 1439 |
| LMArena Creative Writing | 1401 | 1401 |
| EQ-Bench Creative Writing | 1436 | 1703 |
| LMArena Multi-Turn | 1408 | 1458 |
Frequently asked questions
Is DeepSeek-V3.1 better than GPT-5.2?
GPT-5.2 is the stronger model overall, scoring 54.1 to 42.8 on the Noometry Index. DeepSeek-V3.1 costs 11× less per token, which makes it the better buy when GPT-5.2's lead doesn't matter for your workload.
Which is cheaper, DeepSeek-V3.1 or GPT-5.2?
DeepSeek-V3.1 is cheaper. It lists at $0.25 per million input tokens and $0.95 per million output tokens; GPT-5.2 lists at $1.75 and $14.
Is DeepSeek-V3.1 or GPT-5.2 better for coding?
GPT-5.2 scores higher on coding benchmarks: 51.6 versus 40.3 in the Noometry coding category.
Which has the bigger context window?
GPT-5.2 does, with 400K tokens against 164K.
How many benchmarks do DeepSeek-V3.1 and GPT-5.2 share?
26 benchmarks have published results for both models. DeepSeek-V3.1 has 27 scored results on Noometry and GPT-5.2 has 67.