Model comparison
GPT-5.6 Luna vs Phi-4
GPT-5.6 Luna is the stronger model overall, scoring 54.6 to 31.2 on the Noometry Index. Phi-4 costs 5.1× less per token, which makes it the better buy when GPT-5.6 Luna's lead doesn't matter for your workload.
Last verified . 22 shared benchmarks.
MicrosoftSummary
- They share 22 benchmarks with published results for both. GPT-5.6 Luna 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-5.6 Luna leads 77.7 to 20.8.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 98.3% for GPT-5.6 Luna and 13.8% for Phi-4.
- Phi-4 is cheaper at $0.07 / $0.14 per million input/output tokens, against $0.20 / $1.20 for GPT-5.6 Luna.
- GPT-5.6 Luna accepts more context: 1.05M tokens versus 128K.
- Phi-4 has downloadable open weights; the other is API-only.
Side by side
| GPT-5.6 Luna | Phi-4 | |
|---|---|---|
| Provider | OpenAI | Microsoft |
| Noometry Index | 54.6 | 31.2 |
| Released | 2026-07-09 | 2024-12-11 |
| Weights | Proprietary | Open |
| Context window | 1.05M | 128K |
| Max output | 128K | 4K |
| Input $ / M tokens | $0.20 | $0.07 |
| Output $ / M tokens | $1.20 | $0.14 |
| Results tracked | 52 | 37 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding GPT-5.6 Luna leads
GPT-5.6 Luna: 54.5 (#28), Phi-4: 34.4 (#239)
| Benchmark | GPT-5.6 Luna | Phi-4 |
|---|---|---|
| LMArena Coding | 1466 | 1231 |
| DeepSWE | 67.2% | — |
| FrontierCode | 39.8% | — |
| CursorBench | 35.9% | — |
| LMArena WebDev | 1519 | — |
| SciCode | 53.6% | — |
| WeirdML | 60.9% | — |
| BigCodeBench Instruct | — | 45.5% |
| LiveBench Coding | — | 30.7% |
| BigCodeBench Complete | — | 55.4% |
| ALE-Bench | 1,667 | — |
Agentic & Tool Use GPT-5.6 Luna leads
GPT-5.6 Luna: 34.4 (#45), Phi-4: 22.8 (#128)
| Benchmark | GPT-5.6 Luna | Phi-4 |
|---|---|---|
| BALROG | 45.6% | 11.6% |
| APEX-Agents | 43% | — |
| Berkeley Function Calling Leaderboard | — | 28.8% |
| GDP.pdf | 22.7% | — |
| Vending-Bench 2 | 4,095 | — |
Reasoning GPT-5.6 Luna leads
GPT-5.6 Luna: 47.6 (#43), Phi-4: 17.7 (#291)
| Benchmark | GPT-5.6 Luna | Phi-4 |
|---|---|---|
| Chess Puzzles | 40% | 1% |
| LMArena Hard Prompts | 1451 | 1220 |
| Epoch Capabilities Index | 156.39 | 130.42 |
| ARC-AGI-2 | 59.5% | — |
| SimpleBench | 46.8% | — |
| Kagi LLM Benchmark | 49.1% | — |
| NYT Connections (extended) | 69.4% | — |
| ARC-AGI-1 | 88% | — |
| CritPt | 20.6% | — |
| LiveBench Reasoning | — | 47.8% |
| Mystery Game Puzzles | 21% | — |
| DTBench | 89.1% | — |
| LiveBench Data Analysis | — | 45.2% |
| LMCA | 48.5% | — |
| Surface Evolver Bench | 61.9% | — |
| LiveBench | — | 41.6% |
Math GPT-5.6 Luna leads
GPT-5.6 Luna: 77.7 (#14), Phi-4: 20.8 (#285)
| Benchmark | GPT-5.6 Luna | Phi-4 |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 98.3% | 13.8% |
| LMArena Math | 1458 | 1246 |
| FrontierMath (Tiers 1-3) | 82.1% | — |
| FrontierMath Tier 4 | 61% | — |
| ProofBench | 60% | — |
| LiveBench Math | — | 42% |
| MATH Level 5 | — | 64.9% |
Knowledge GPT-5.6 Luna leads
GPT-5.6 Luna: 58.5 (#34), Phi-4: 32.6 (#209)
| Benchmark | GPT-5.6 Luna | Phi-4 |
|---|---|---|
| GPQA Diamond | 91.6% | 56.1% |
| LMArena Expert | 1478 | 1203 |
| SimpleQA Verified | 41% | — |
| Confabulations | — | 29.4% |
| Vectara Hallucination Rate | — | 3.7% |
| MMLU | — | 84.8% |
Multimodal Not comparable
GPT-5.6 Luna: 42.7 (#28), Phi-4: —
| Benchmark | GPT-5.6 Luna | Phi-4 |
|---|---|---|
| LMArena Vision | 1258 | — |
| Blueprint-Bench 2 | 22.6% | — |
| Furniture Assembly | 42.5% | — |
| LMArena Document | 1457 | — |
Multilingual GPT-5.6 Luna leads
GPT-5.6 Luna: 52.8 (#78), Phi-4: 37.2 (#237)
| Benchmark | GPT-5.6 Luna | Phi-4 |
|---|---|---|
| LMArena Non-English | 1417 | 1197 |
| LMArena Chinese | 1470 | 1212 |
| LMArena French | 1456 | 1224 |
| LMArena German | 1454 | 1222 |
| LMArena Japanese | 1411 | 1158 |
| LMArena Korean | 1415 | 1151 |
| LMArena Russian | 1428 | 1209 |
| LMArena Spanish | 1448 | 1234 |
Instruction Following GPT-5.6 Luna leads
GPT-5.6 Luna: 75.6 (#57), Phi-4: 60.4 (#251)
| Benchmark | GPT-5.6 Luna | Phi-4 |
|---|---|---|
| LMArena Instruction Following | 1437 | 1201 |
| LiveBench Instruction Following | — | 58.4% |
Long Context GPT-5.6 Luna leads
GPT-5.6 Luna: 43.9 (#82), Phi-4: 36.9 (#226)
| Benchmark | GPT-5.6 Luna | Phi-4 |
|---|---|---|
| LMArena Longer Query | 1436 | 1217 |
Writing & Preference GPT-5.6 Luna leads
GPT-5.6 Luna: 68.0 (#29), Phi-4: 40.5 (#244)
| Benchmark | GPT-5.6 Luna | Phi-4 |
|---|---|---|
| LMArena Text | 1431 | 1217 |
| LMArena Creative Writing | 1396 | 1182 |
| LMArena Multi-Turn | 1434 | 1206 |
| Short-Story Creative Writing | — | 62.6% |
| EQ-Bench Creative Writing | 1829 | — |
| EQ-Bench 4 | 1156 | — |
| LiveBench Language | — | 25.6% |
Frequently asked questions
Is GPT-5.6 Luna better than Phi-4?
GPT-5.6 Luna is the stronger model overall, scoring 54.6 to 31.2 on the Noometry Index. Phi-4 costs 5.1× less per token, which makes it the better buy when GPT-5.6 Luna's lead doesn't matter for your workload.
Which is cheaper, GPT-5.6 Luna or Phi-4?
Phi-4 is cheaper. It lists at $0.07 per million input tokens and $0.14 per million output tokens; GPT-5.6 Luna lists at $0.20 and $1.20.
Is GPT-5.6 Luna or Phi-4 better for coding?
GPT-5.6 Luna scores higher on coding benchmarks: 54.5 versus 34.4 in the Noometry coding category.
Which has the bigger context window?
GPT-5.6 Luna does, with 1.05M tokens against 128K.
How many benchmarks do GPT-5.6 Luna and Phi-4 share?
22 benchmarks have published results for both models. GPT-5.6 Luna has 52 scored results on Noometry and Phi-4 has 37.