Model comparison
DeepSeek-R1 vs DeepSeek-V3
DeepSeek-R1 is the stronger model overall, scoring 42.3 to 39.5 on the Noometry Index. DeepSeek-V3 costs 2.3× less per token, which makes it the better buy when DeepSeek-R1's lead doesn't matter for your workload.
Last verified . 46 shared benchmarks.
DeepSeekSummary
- They share 46 benchmarks with published results for both. DeepSeek-R1 scores higher in 6 categories and DeepSeek-V3 in 2 categories; 7 gaps are clear of the uncertainty.
- The widest gap is in math, where DeepSeek-R1 leads 43.8 to 32.1.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 66.4% for DeepSeek-R1 and 37.8% for DeepSeek-V3.
- DeepSeek-V3 is cheaper at $0.24 / $0.90 per million input/output tokens, against $0.50 / $2.15 for DeepSeek-R1.
- DeepSeek-V3 has downloadable open weights; the other is API-only.
Side by side
| DeepSeek-R1 | DeepSeek-V3 | |
|---|---|---|
| Provider | DeepSeek | DeepSeek |
| Noometry Index | 42.3 | 39.5 |
| Released | 2025-01-20 | 2024-12-26 |
| Weights | Proprietary | Open |
| Context window | 164K | 164K |
| Max output | 64K | 164K |
| Input $ / M tokens | $0.50 | $0.24 |
| Output $ / M tokens | $2.15 | $0.90 |
| Results tracked | 52 | 60 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek-R1 leads
DeepSeek-R1: 46.3 (#68), DeepSeek-V3: 42.3 (#106)
| Benchmark | DeepSeek-R1 | DeepSeek-V3 |
|---|---|---|
| Aider Polyglot | 71.4% | 55.1% |
| SciCode | 35.7% | 35.8% |
| WeirdML | 41.6% | 36.1% |
| LiveBench Coding | 66.7% | 70.9% |
| LMArena Coding | 1427 | 1368 |
| BigCodeBench Instruct | — | 50% |
| BigCodeBench Complete | — | 62.2% |
| ALE-Bench | 804.12 | — |
| AlgoTune | 1.7 | — |
| HumanEval+ | — | 86.6% |
| MBPP+ | — | 73% |
Agentic & Tool Use Not comparable
DeepSeek-R1: 30.7 (#75), DeepSeek-V3: —
| Benchmark | DeepSeek-R1 | DeepSeek-V3 |
|---|---|---|
| METR Time Horizons | 53.8% | 49.6% |
| DeepResearch Bench | 35.1% | — |
| BALROG | 34.9% | — |
Reasoning DeepSeek-V3 leads
DeepSeek-R1: 18.6 (#278), DeepSeek-V3: 20.5 (#236)
| Benchmark | DeepSeek-R1 | DeepSeek-V3 |
|---|---|---|
| SimpleBench | 40.8% | 27.2% |
| Kagi LLM Benchmark | 69.4% | 52.3% |
| CritPt | 1.1% | 0% |
| LiveBench Reasoning | 83.2% | 65.8% |
| LMArena Hard Prompts | 1416 | 1365 |
| LiveBench Data Analysis | 69.8% | 60.9% |
| Epoch Capabilities Index | 141.29 | 135.94 |
| ForecastBench | 60 | 59.1 |
| LiveBench | 71.6% | 66.9% |
| ARC-AGI-2 | 1.3% | — |
| ARC-AGI-1 | 21.2% | — |
| DTBench | — | 64.8% |
| LMCA | — | 15.5% |
| BIG-Bench Hard | — | 87.5% |
| HellaSwag | — | 88.9% |
| PIQA | — | 84.7% |
| WinoGrande | — | 85.2% |
Math DeepSeek-R1 leads
DeepSeek-R1: 43.8 (#79), DeepSeek-V3: 32.1 (#219)
| Benchmark | DeepSeek-R1 | DeepSeek-V3 |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 66.4% | 37.8% |
| Omni-MATH | 42.4% | 40.3% |
| LiveBench Math | 80.7% | 73.5% |
| LMArena Math | 1400 | 1373 |
| MATH Level 5 | 96.6% | 75.5% |
| FrontierMath (Feb 2025 set) | — | 1.7% |
Knowledge DeepSeek-R1 leads
DeepSeek-R1: 44.5 (#87), DeepSeek-V3: 37.5 (#155)
| Benchmark | DeepSeek-R1 | DeepSeek-V3 |
|---|---|---|
| GPQA Diamond | 76.3% | 67.6% |
| MMLU-Pro | 79.3% | 72.3% |
| Confabulations | 12.7% | 26.1% |
| Vectara Hallucination Rate | 11.3% | 6.1% |
| GPQA (HELM) | 66.6% | 53.8% |
| LMArena Expert | 1394 | 1351 |
| ARC (AI2) Challenge | — | 95.3% |
| MMLU | — | 87.2% |
| TriviaQA | — | 82.9% |
Multilingual DeepSeek-R1 leads
DeepSeek-R1: 52.4 (#85), DeepSeek-V3: 48.5 (#143)
| Benchmark | DeepSeek-R1 | DeepSeek-V3 |
|---|---|---|
| LMArena Non-English | 1412 | 1358 |
| LMArena Chinese | 1442 | 1391 |
| LMArena French | 1417 | 1385 |
| LMArena German | 1404 | 1374 |
| LMArena Japanese | 1391 | 1333 |
| LMArena Korean | 1360 | 1319 |
| LMArena Russian | 1423 | 1373 |
| LMArena Spanish | 1411 | 1358 |
Instruction Following Too close to call
DeepSeek-R1: 72.0 (#143), DeepSeek-V3: 72.8 (#130)
| Benchmark | DeepSeek-R1 | DeepSeek-V3 |
|---|---|---|
| LiveBench Instruction Following | 80.5% | 81.5% |
| IFEval | 78.4% | 83.2% |
| LMArena Instruction Following | 1382 | 1345 |
Long Context DeepSeek-R1 leads
DeepSeek-R1: 45.4 (#36), DeepSeek-V3: 34.0 (#253)
| Benchmark | DeepSeek-R1 | DeepSeek-V3 |
|---|---|---|
| Fiction.LiveBench | 75% | 50% |
| LMArena Longer Query | 1391 | 1352 |
Writing & Preference DeepSeek-R1 leads
DeepSeek-R1: 61.4 (#88), DeepSeek-V3: 57.4 (#130)
| Benchmark | DeepSeek-R1 | DeepSeek-V3 |
|---|---|---|
| LMArena Text | 1428 | 1375 |
| LMArena Creative Writing | 1405 | 1364 |
| Short-Story Creative Writing | 83% | 77% |
| EQ-Bench Creative Writing | 1500 | 1472 |
| WildBench | 82.8% | 83% |
| LMArena Multi-Turn | 1405 | 1389 |
| LiveBench Language | 48.5% | 49.1% |
Frequently asked questions
Is DeepSeek-R1 better than DeepSeek-V3?
DeepSeek-R1 is the stronger model overall, scoring 42.3 to 39.5 on the Noometry Index. DeepSeek-V3 costs 2.3× less per token, which makes it the better buy when DeepSeek-R1's lead doesn't matter for your workload.
Which is cheaper, DeepSeek-R1 or DeepSeek-V3?
DeepSeek-V3 is cheaper. It lists at $0.24 per million input tokens and $0.90 per million output tokens; DeepSeek-R1 lists at $0.50 and $2.15.
Is DeepSeek-R1 or DeepSeek-V3 better for coding?
DeepSeek-R1 scores higher on coding benchmarks: 46.3 versus 42.3 in the Noometry coding category.
Which has the bigger context window?
Both accept 164K tokens.
How many benchmarks do DeepSeek-R1 and DeepSeek-V3 share?
46 benchmarks have published results for both models. DeepSeek-R1 has 52 scored results on Noometry and DeepSeek-V3 has 60.