Model comparison
GPT-5 Nano vs Phi-4
GPT-5 Nano is the stronger model overall, scoring 33.5 to 31.2 on the Noometry Index. Phi-4 costs 1.6× less per token, which makes it the better buy when GPT-5 Nano's lead doesn't matter for your workload.
Last verified . 23 shared benchmarks.
MicrosoftSummary
- They share 23 benchmarks with published results for both. GPT-5 Nano scores higher in 5 categories and Phi-4 in 4 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in instruction following, where GPT-5 Nano leads 75.0 to 60.4.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 81.1% for GPT-5 Nano and 13.8% for Phi-4.
- Phi-4 is cheaper at $0.07 / $0.14 per million input/output tokens, against $0.05 / $0.40 for GPT-5 Nano.
- GPT-5 Nano accepts more context: 400K tokens versus 128K.
- Phi-4 has downloadable open weights; the other is API-only.
Side by side
| GPT-5 Nano | Phi-4 | |
|---|---|---|
| Provider | OpenAI | Microsoft |
| Noometry Index | 33.5 | 31.2 |
| Released | 2025-08-07 | 2024-12-11 |
| Weights | Proprietary | Open |
| Context window | 400K | 128K |
| Max output | 128K | 4K |
| Input $ / M tokens | $0.05 | $0.07 |
| Output $ / M tokens | $0.40 | $0.14 |
| Results tracked | 49 | 37 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding Too close to call
GPT-5 Nano: 33.6 (#254), Phi-4: 34.4 (#239)
| Benchmark | GPT-5 Nano | Phi-4 |
|---|---|---|
| LMArena Coding | 1351 | 1231 |
| SWE-bench Verified (bash only) | 34.8% | — |
| WeirdML | 38.1% | — |
| BigCodeBench Instruct | — | 45.5% |
| LiveBench Coding | — | 30.7% |
| BigCodeBench Complete | — | 55.4% |
| ALE-Bench | 718.67 | — |
Agentic & Tool Use GPT-5 Nano leads
GPT-5 Nano: 25.8 (#106), Phi-4: 22.8 (#128)
| Benchmark | GPT-5 Nano | Phi-4 |
|---|---|---|
| Berkeley Function Calling Leaderboard | 51.5% | 28.8% |
| Terminal-Bench | 21.8% | — |
| BALROG | — | 11.6% |
Reasoning Phi-4 leads
GPT-5 Nano: 16.3 (#306), Phi-4: 17.7 (#291)
| Benchmark | GPT-5 Nano | Phi-4 |
|---|---|---|
| Chess Puzzles | 27% | 1% |
| LMArena Hard Prompts | 1328 | 1220 |
| Epoch Capabilities Index | 139.38 | 130.42 |
| ARC-AGI-2 | 2.6% | — |
| Kagi LLM Benchmark | 62.2% | — |
| ARC-AGI-1 | 20.7% | — |
| LiveBench Reasoning | — | 47.8% |
| Mystery Game Puzzles | 9% | — |
| DTBench | 62.7% | — |
| LiveBench Data Analysis | — | 45.2% |
| LMCA | 7.9% | — |
| ForecastBench | 59.1 | — |
| LiveBench | — | 41.6% |
Math GPT-5 Nano leads
GPT-5 Nano: 29.4 (#241), Phi-4: 20.8 (#285)
| Benchmark | GPT-5 Nano | Phi-4 |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 81.1% | 13.8% |
| LMArena Math | 1317 | 1246 |
| MATH Level 5 | 95.2% | 64.9% |
| FrontierMath (Tiers 1-3) | 20% | — |
| FrontierMath Tier 4 | 2.4% | — |
| ProofBench | 12% | — |
| Omni-MATH | 54.6% | — |
| LiveBench Math | — | 42% |
| FrontierMath (Feb 2025 set) | 8.3% | — |
| FrontierMath Tier 4 (v1) | 2.1% | — |
Knowledge GPT-5 Nano leads
GPT-5 Nano: 35.9 (#178), Phi-4: 32.6 (#209)
| Benchmark | GPT-5 Nano | Phi-4 |
|---|---|---|
| GPQA Diamond | 69.4% | 56.1% |
| Vectara Hallucination Rate | 10.5% | 3.7% |
| LMArena Expert | 1321 | 1203 |
| SimpleQA Verified | 11.7% | — |
| MMLU-Pro | 77.8% | — |
| Confabulations | — | 29.4% |
| GPQA (HELM) | 67.9% | — |
| MMLU | — | 84.8% |
Multimodal Not comparable
GPT-5 Nano: 31.3 (#108), Phi-4: —
| Benchmark | GPT-5 Nano | Phi-4 |
|---|---|---|
| LMArena Vision | 1159 | — |
| VPCT | 37.2% | — |
Multilingual GPT-5 Nano leads
GPT-5 Nano: 45.3 (#172), Phi-4: 37.2 (#237)
| Benchmark | GPT-5 Nano | Phi-4 |
|---|---|---|
| LMArena Non-English | 1313 | 1197 |
| LMArena Chinese | 1356 | 1212 |
| LMArena German | 1327 | 1222 |
| LMArena Japanese | 1226 | 1158 |
| LMArena Korean | 1269 | 1151 |
| LMArena Russian | 1296 | 1209 |
| LMArena Spanish | 1360 | 1234 |
| LMArena French | — | 1224 |
Instruction Following GPT-5 Nano leads
GPT-5 Nano: 75.0 (#79), Phi-4: 60.4 (#251)
| Benchmark | GPT-5 Nano | Phi-4 |
|---|---|---|
| LMArena Instruction Following | 1306 | 1201 |
| LiveBench Instruction Following | — | 58.4% |
| IFEval | 93.2% | — |
Long Context Phi-4 leads
GPT-5 Nano: 31.3 (#281), Phi-4: 36.9 (#226)
| Benchmark | GPT-5 Nano | Phi-4 |
|---|---|---|
| LMArena Longer Query | 1312 | 1217 |
| Fiction.LiveBench | 44.4% | — |
Writing & Preference Phi-4 leads
GPT-5 Nano: 39.1 (#249), Phi-4: 40.5 (#244)
| Benchmark | GPT-5 Nano | Phi-4 |
|---|---|---|
| LMArena Text | 1320 | 1217 |
| LMArena Creative Writing | 1249 | 1182 |
| LMArena Multi-Turn | 1311 | 1206 |
| Short-Story Creative Writing | — | 62.6% |
| EQ-Bench Creative Writing | 705 | — |
| WildBench | 80.6% | — |
| LiveBench Language | — | 25.6% |
Frequently asked questions
Is GPT-5 Nano better than Phi-4?
GPT-5 Nano is the stronger model overall, scoring 33.5 to 31.2 on the Noometry Index. Phi-4 costs 1.6× less per token, which makes it the better buy when GPT-5 Nano's lead doesn't matter for your workload.
Which is cheaper, GPT-5 Nano or Phi-4?
Phi-4 is cheaper. It lists at $0.07 per million input tokens and $0.14 per million output tokens; GPT-5 Nano lists at $0.05 and $0.40.
Is GPT-5 Nano or Phi-4 better for coding?
They score almost the same on coding (33.6 vs 34.4); test both on your own repository before choosing.
Which has the bigger context window?
GPT-5 Nano does, with 400K tokens against 128K.
How many benchmarks do GPT-5 Nano and Phi-4 share?
23 benchmarks have published results for both models. GPT-5 Nano has 49 scored results on Noometry and Phi-4 has 37.