Model comparison
DeepSeek-R1 vs Phi-4
DeepSeek-R1 is the stronger model overall, scoring 42.3 to 31.2 on the Noometry Index. Phi-4 costs 10× less per token, which makes it the better buy when DeepSeek-R1's lead doesn't matter for your workload.
Last verified . 32 shared benchmarks.
DeepSeek
MicrosoftSummary
- They share 32 benchmarks with published results for both. DeepSeek-R1 scores higher in 9 categories and Phi-4 in 0 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in math, where DeepSeek-R1 leads 43.8 to 20.8.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 66.4% for DeepSeek-R1 and 13.8% for Phi-4.
- Phi-4 is cheaper at $0.07 / $0.14 per million input/output tokens, against $0.50 / $2.15 for DeepSeek-R1.
- DeepSeek-R1 accepts more context: 164K tokens versus 128K.
- Phi-4 has downloadable open weights; the other is API-only.
Side by side
| DeepSeek-R1 | Phi-4 | |
|---|---|---|
| Provider | DeepSeek | Microsoft |
| Noometry Index | 42.3 | 31.2 |
| Released | 2025-01-20 | 2024-12-11 |
| Weights | Proprietary | Open |
| Context window | 164K | 128K |
| Max output | 64K | 4K |
| Input $ / M tokens | $0.50 | $0.07 |
| Output $ / M tokens | $2.15 | $0.14 |
| Results tracked | 52 | 37 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek-R1 leads
DeepSeek-R1: 46.3 (#68), Phi-4: 34.4 (#239)
| Benchmark | DeepSeek-R1 | Phi-4 |
|---|---|---|
| LiveBench Coding | 66.7% | 30.7% |
| LMArena Coding | 1427 | 1231 |
| Aider Polyglot | 71.4% | — |
| SciCode | 35.7% | — |
| WeirdML | 41.6% | — |
| BigCodeBench Instruct | — | 45.5% |
| BigCodeBench Complete | — | 55.4% |
| ALE-Bench | 804.12 | — |
| AlgoTune | 1.7 | — |
Agentic & Tool Use DeepSeek-R1 leads
DeepSeek-R1: 30.7 (#75), Phi-4: 22.8 (#128)
| Benchmark | DeepSeek-R1 | Phi-4 |
|---|---|---|
| BALROG | 34.9% | 11.6% |
| Berkeley Function Calling Leaderboard | — | 28.8% |
| DeepResearch Bench | 35.1% | — |
| METR Time Horizons | 53.8% | — |
Reasoning Too close to call
DeepSeek-R1: 18.6 (#278), Phi-4: 17.7 (#291)
| Benchmark | DeepSeek-R1 | Phi-4 |
|---|---|---|
| LiveBench Reasoning | 83.2% | 47.8% |
| LMArena Hard Prompts | 1416 | 1220 |
| LiveBench Data Analysis | 69.8% | 45.2% |
| Epoch Capabilities Index | 141.29 | 130.42 |
| LiveBench | 71.6% | 41.6% |
| ARC-AGI-2 | 1.3% | — |
| SimpleBench | 40.8% | — |
| Kagi LLM Benchmark | 69.4% | — |
| ARC-AGI-1 | 21.2% | — |
| CritPt | 1.1% | — |
| Chess Puzzles | — | 1% |
| ForecastBench | 60 | — |
Math DeepSeek-R1 leads
DeepSeek-R1: 43.8 (#79), Phi-4: 20.8 (#285)
| Benchmark | DeepSeek-R1 | Phi-4 |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 66.4% | 13.8% |
| LiveBench Math | 80.7% | 42% |
| LMArena Math | 1400 | 1246 |
| MATH Level 5 | 96.6% | 64.9% |
| Omni-MATH | 42.4% | — |
Knowledge DeepSeek-R1 leads
DeepSeek-R1: 44.5 (#87), Phi-4: 32.6 (#209)
| Benchmark | DeepSeek-R1 | Phi-4 |
|---|---|---|
| GPQA Diamond | 76.3% | 56.1% |
| Confabulations | 12.7% | 29.4% |
| Vectara Hallucination Rate | 11.3% | 3.7% |
| LMArena Expert | 1394 | 1203 |
| MMLU-Pro | 79.3% | — |
| GPQA (HELM) | 66.6% | — |
| MMLU | — | 84.8% |
Multilingual DeepSeek-R1 leads
DeepSeek-R1: 52.4 (#85), Phi-4: 37.2 (#237)
| Benchmark | DeepSeek-R1 | Phi-4 |
|---|---|---|
| LMArena Non-English | 1412 | 1197 |
| LMArena Chinese | 1442 | 1212 |
| LMArena French | 1417 | 1224 |
| LMArena German | 1404 | 1222 |
| LMArena Japanese | 1391 | 1158 |
| LMArena Korean | 1360 | 1151 |
| LMArena Russian | 1423 | 1209 |
| LMArena Spanish | 1411 | 1234 |
Instruction Following DeepSeek-R1 leads
DeepSeek-R1: 72.0 (#143), Phi-4: 60.4 (#251)
| Benchmark | DeepSeek-R1 | Phi-4 |
|---|---|---|
| LiveBench Instruction Following | 80.5% | 58.4% |
| LMArena Instruction Following | 1382 | 1201 |
| IFEval | 78.4% | — |
Long Context DeepSeek-R1 leads
DeepSeek-R1: 45.4 (#36), Phi-4: 36.9 (#226)
| Benchmark | DeepSeek-R1 | Phi-4 |
|---|---|---|
| LMArena Longer Query | 1391 | 1217 |
| Fiction.LiveBench | 75% | — |
Writing & Preference DeepSeek-R1 leads
DeepSeek-R1: 61.4 (#88), Phi-4: 40.5 (#244)
| Benchmark | DeepSeek-R1 | Phi-4 |
|---|---|---|
| LMArena Text | 1428 | 1217 |
| LMArena Creative Writing | 1405 | 1182 |
| Short-Story Creative Writing | 83% | 62.6% |
| LMArena Multi-Turn | 1405 | 1206 |
| LiveBench Language | 48.5% | 25.6% |
| EQ-Bench Creative Writing | 1500 | — |
| WildBench | 82.8% | — |
Frequently asked questions
Is DeepSeek-R1 better than Phi-4?
DeepSeek-R1 is the stronger model overall, scoring 42.3 to 31.2 on the Noometry Index. Phi-4 costs 10× 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 Phi-4?
Phi-4 is cheaper. It lists at $0.07 per million input tokens and $0.14 per million output tokens; DeepSeek-R1 lists at $0.50 and $2.15.
Is DeepSeek-R1 or Phi-4 better for coding?
DeepSeek-R1 scores higher on coding benchmarks: 46.3 versus 34.4 in the Noometry coding category.
Which has the bigger context window?
DeepSeek-R1 does, with 164K tokens against 128K.
How many benchmarks do DeepSeek-R1 and Phi-4 share?
32 benchmarks have published results for both models. DeepSeek-R1 has 52 scored results on Noometry and Phi-4 has 37.