# GPT-5 Mini vs Phi-4

> GPT-5 Mini is the stronger model overall, scoring 41.8 to 31.2 on the Noometry Index. Phi-4 costs 7.9× less per token, which makes it the better buy when GPT-5 Mini's lead doesn't matter for your workload.

- Canonical page: https://noometry.com/compare/gpt-5-mini-vs-phi-4
- Last updated: 2026-10-11
- Shared benchmarks: 26

## Summary

- They share 26 benchmarks with published results for both. GPT-5 Mini scores higher in 9 categories and Phi-4 in 0 categories; 9 gaps are clear of the uncertainty.
- The widest gap is in math, where GPT-5 Mini leads 46.7 to 20.8.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 86.7% for GPT-5 Mini and 13.8% for Phi-4.
- Phi-4 is cheaper at $0.07 / $0.14 per million input/output tokens, against $0.25 / $2 for GPT-5 Mini.
- GPT-5 Mini accepts more context: 400K tokens versus 128K.
- Phi-4 has downloadable open weights; the other is API-only.

## Snapshot

| | GPT-5 Mini | Phi-4 |
|---|---|---|
| Provider | OpenAI | Microsoft |
| Noometry Index | 41.8 | 31.2 |
| Rank | 128 | 279 |
| Context | 400K | 128K |
| Input $/M | $0.25 | $0.07 |
| Output $/M | $2 | $0.14 |
| Weights | Proprietary | Open |

## Coding

- GPT-5 Mini: 40.1 (#146)
- Phi-4: 34.4 (#239)

| Benchmark | GPT-5 Mini | Phi-4 |
|---|---|---|
| LMArena Coding | 1406 | 1231 |
| SWE-bench Verified | 64.7% | — |
| SWE-bench Verified (bash only) | 59.8% | — |
| SWE-bench Multilingual | 39.7% | — |
| SciCode | 39.2% | — |
| WeirdML | 52.7% | — |
| BigCodeBench Instruct | — | 45.5% |
| LiveBench Coding | — | 30.7% |
| BigCodeBench Complete | — | 55.4% |
| ALE-Bench | 799.77 | — |
| AlgoTune | 1.38 | — |

## Agentic & Tool Use

- GPT-5 Mini: 31.1 (#70)
- Phi-4: 22.8 (#128)

| Benchmark | GPT-5 Mini | Phi-4 |
|---|---|---|
| Berkeley Function Calling Leaderboard | 55.5% | 28.8% |
| Terminal-Bench | 34.8% | — |
| BALROG | — | 11.6% |
| Vending-Bench 2 | -31.18 | — |

## Reasoning

- GPT-5 Mini: 23.9 (#168)
- Phi-4: 17.7 (#291)

| Benchmark | GPT-5 Mini | Phi-4 |
|---|---|---|
| Chess Puzzles | 30% | 1% |
| LMArena Hard Prompts | 1380 | 1220 |
| Epoch Capabilities Index | 145.52 | 130.42 |
| ARC-AGI-2 | 4.4% | — |
| Kagi LLM Benchmark | 70.3% | — |
| ARC-AGI-1 | 54.3% | — |
| CritPt | 0% | — |
| EnigmaEval | 8.2% | — |
| LiveBench Reasoning | — | 47.8% |
| Mystery Game Puzzles | 10% | — |
| DTBench | 80.5% | — |
| LiveBench Data Analysis | — | 45.2% |
| LMCA | 34.2% | — |
| ForecastBench | 61 | — |
| LiveBench | — | 41.6% |

## Math

- GPT-5 Mini: 46.7 (#69)
- Phi-4: 20.8 (#285)

| Benchmark | GPT-5 Mini | Phi-4 |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 86.7% | 13.8% |
| LMArena Math | 1378 | 1246 |
| MATH Level 5 | 97.8% | 64.9% |
| FrontierMath (Tiers 1-3) | 46.7% | — |
| FrontierMath Tier 4 | 12.2% | — |
| ProofBench | 9% | — |
| Omni-MATH | 72.2% | — |
| LiveBench Math | — | 42% |
| FrontierMath (Feb 2025 set) | 27.2% | — |
| FrontierMath Tier 4 (v1) | 6.3% | — |

## Knowledge

- GPT-5 Mini: 45.6 (#86)
- Phi-4: 32.6 (#209)

| Benchmark | GPT-5 Mini | Phi-4 |
|---|---|---|
| GPQA Diamond | 75% | 56.1% |
| Confabulations | 13.3% | 29.4% |
| Vectara Hallucination Rate | 12.9% | 3.7% |
| LMArena Expert | 1379 | 1203 |
| Humanity's Last Exam | 19.4% | — |
| SimpleQA Verified | 21.6% | — |
| MMLU-Pro | 83.5% | — |
| GPQA (HELM) | 75.6% | — |
| MMLU | — | 84.8% |

## Multimodal

- GPT-5 Mini: 35.6 (#85)
- Phi-4: —

| Benchmark | GPT-5 Mini | Phi-4 |
|---|---|---|
| LMArena Vision | 1202 | — |
| VPCT | 40.2% | — |

## Multilingual

- GPT-5 Mini: 48.9 (#137)
- Phi-4: 37.2 (#237)

| Benchmark | GPT-5 Mini | Phi-4 |
|---|---|---|
| LMArena Non-English | 1363 | 1197 |
| LMArena Chinese | 1385 | 1212 |
| LMArena French | 1386 | 1224 |
| LMArena German | 1366 | 1222 |
| LMArena Japanese | 1341 | 1158 |
| LMArena Korean | 1308 | 1151 |
| LMArena Russian | 1362 | 1209 |
| LMArena Spanish | 1355 | 1234 |

## Instruction Following

- GPT-5 Mini: 76.2 (#46)
- Phi-4: 60.4 (#251)

| Benchmark | GPT-5 Mini | Phi-4 |
|---|---|---|
| LMArena Instruction Following | 1357 | 1201 |
| LiveBench Instruction Following | — | 58.4% |
| IFEval | 92.7% | — |

## Long Context

- GPT-5 Mini: 41.9 (#132)
- Phi-4: 36.9 (#226)

| Benchmark | GPT-5 Mini | Phi-4 |
|---|---|---|
| LMArena Longer Query | 1355 | 1217 |
| Fiction.LiveBench | 69.4% | — |

## Writing & Preference

- GPT-5 Mini: 55.2 (#148)
- Phi-4: 40.5 (#244)

| Benchmark | GPT-5 Mini | Phi-4 |
|---|---|---|
| LMArena Text | 1373 | 1217 |
| LMArena Creative Writing | 1325 | 1182 |
| Short-Story Creative Writing | 83.1% | 62.6% |
| LMArena Multi-Turn | 1363 | 1206 |
| EQ-Bench Creative Writing | 1313 | — |
| WildBench | 85.5% | — |
| LiveBench Language | — | 25.6% |

## FAQ

### Is GPT-5 Mini better than Phi-4?

GPT-5 Mini is the stronger model overall, scoring 41.8 to 31.2 on the Noometry Index. Phi-4 costs 7.9× less per token, which makes it the better buy when GPT-5 Mini's lead doesn't matter for your workload.

### Which is cheaper, GPT-5 Mini 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 Mini lists at $0.25 and $2.

### Is GPT-5 Mini or Phi-4 better for coding?

GPT-5 Mini scores higher on coding benchmarks: 40.1 versus 34.4 in the Noometry coding category.

### Which has the bigger context window?

GPT-5 Mini does, with 400K tokens against 128K.

### How many benchmarks do GPT-5 Mini and Phi-4 share?

26 benchmarks have published results for both models. GPT-5 Mini has 60 scored results on Noometry and Phi-4 has 37.