Model comparison
GPT-5.4 mini vs GPT-5.6 Luna
GPT-5.6 Luna is the stronger model overall, scoring 54.6 to 45.0 on the Noometry Index.
Last verified . 40 shared benchmarks.
Summary
- They share 40 benchmarks with published results for both. GPT-5.4 mini scores higher in 0 categories and GPT-5.6 Luna in 10 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in math, where GPT-5.6 Luna leads 77.7 to 45.5.
- The biggest single-benchmark swing is FrontierMath Tier 4: 9.8% for GPT-5.4 mini and 61% for GPT-5.6 Luna.
- GPT-5.6 Luna is cheaper at $0.20 / $1.20 per million input/output tokens, against $0.75 / $4.50 for GPT-5.4 mini.
- GPT-5.6 Luna accepts more context: 1.05M tokens versus 400K.
Side by side
| GPT-5.4 mini | GPT-5.6 Luna | |
|---|---|---|
| Provider | OpenAI | OpenAI |
| Noometry Index | 45.0 | 54.6 |
| Released | 2026-03-17 | 2026-07-09 |
| Weights | Proprietary | Proprietary |
| Context window | 400K | 1.05M |
| Max output | 128K | 128K |
| Input $ / M tokens | $0.75 | $0.20 |
| Output $ / M tokens | $4.50 | $1.20 |
| Results tracked | 46 | 52 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding GPT-5.6 Luna leads
GPT-5.4 mini: 45.2 (#72), GPT-5.6 Luna: 54.5 (#28)
| Benchmark | GPT-5.4 mini | GPT-5.6 Luna |
|---|---|---|
| FrontierCode | 27% | 39.8% |
| LMArena WebDev | 1397 | 1519 |
| SciCode | 49.9% | 53.6% |
| WeirdML | 60.3% | 60.9% |
| LMArena Coding | 1438 | 1466 |
| ALE-Bench | 1,189 | 1,667 |
| DeepSWE | — | 67.2% |
| CursorBench | — | 35.9% |
Agentic & Tool Use GPT-5.6 Luna leads
GPT-5.4 mini: 29.9 (#81), GPT-5.6 Luna: 34.4 (#45)
| Benchmark | GPT-5.4 mini | GPT-5.6 Luna |
|---|---|---|
| APEX-Agents | — | 43% |
| DeepResearch Bench | 36.3% | — |
| BALROG | — | 45.6% |
| GDP.pdf | — | 22.7% |
| Vending-Bench 2 | — | 4,095 |
Reasoning GPT-5.6 Luna leads
GPT-5.4 mini: 30.4 (#85), GPT-5.6 Luna: 47.6 (#43)
| Benchmark | GPT-5.4 mini | GPT-5.6 Luna |
|---|---|---|
| ARC-AGI-2 | 18.9% | 59.5% |
| Kagi LLM Benchmark | 37.9% | 49.1% |
| NYT Connections (extended) | 61.8% | 69.4% |
| ARC-AGI-1 | 63.7% | 88% |
| CritPt | 10% | 20.6% |
| Chess Puzzles | 24% | 40% |
| LMArena Hard Prompts | 1424 | 1451 |
| Mystery Game Puzzles | 11% | 21% |
| DTBench | 80% | 89.1% |
| LMCA | 40.8% | 48.5% |
| Epoch Capabilities Index | 148.84 | 156.39 |
| SimpleBench | — | 46.8% |
| Thematic Generalization | 61.7% | — |
| Surface Evolver Bench | — | 61.9% |
| ForecastBench | 57 | — |
Math GPT-5.6 Luna leads
GPT-5.4 mini: 45.5 (#75), GPT-5.6 Luna: 77.7 (#14)
| Benchmark | GPT-5.4 mini | GPT-5.6 Luna |
|---|---|---|
| FrontierMath (Tiers 1-3) | 51.2% | 82.1% |
| FrontierMath Tier 4 | 9.8% | 61% |
| OTIS Mock AIME 2024-2025 | 88.9% | 98.3% |
| ProofBench | 21% | 60% |
| LMArena Math | 1419 | 1458 |
| FrontierMath (Feb 2025 set) | 28.3% | — |
| FrontierMath Tier 4 (v1) | 2.1% | — |
Knowledge GPT-5.6 Luna leads
GPT-5.4 mini: 51.5 (#67), GPT-5.6 Luna: 58.5 (#34)
| Benchmark | GPT-5.4 mini | GPT-5.6 Luna |
|---|---|---|
| GPQA Diamond | 86.9% | 91.6% |
| SimpleQA Verified | 29.4% | 41% |
| LMArena Expert | 1435 | 1478 |
| Vectara Hallucination Rate | 5.5% | — |
Multimodal GPT-5.6 Luna leads
GPT-5.4 mini: 39.7 (#56), GPT-5.6 Luna: 42.7 (#28)
| Benchmark | GPT-5.4 mini | GPT-5.6 Luna |
|---|---|---|
| LMArena Vision | 1245 | 1258 |
| Blueprint-Bench 2 | — | 22.6% |
| Furniture Assembly | — | 42.5% |
| LMArena Document | — | 1457 |
Multilingual Too close to call
GPT-5.4 mini: 51.9 (#96), GPT-5.6 Luna: 52.8 (#78)
| Benchmark | GPT-5.4 mini | GPT-5.6 Luna |
|---|---|---|
| LMArena Non-English | 1405 | 1417 |
| LMArena Chinese | 1446 | 1470 |
| LMArena French | 1440 | 1456 |
| LMArena German | 1409 | 1454 |
| LMArena Japanese | 1374 | 1411 |
| LMArena Korean | 1368 | 1415 |
| LMArena Russian | 1417 | 1428 |
| LMArena Spanish | 1405 | 1448 |
Instruction Following GPT-5.6 Luna leads
GPT-5.4 mini: 74.1 (#102), GPT-5.6 Luna: 75.6 (#57)
| Benchmark | GPT-5.4 mini | GPT-5.6 Luna |
|---|---|---|
| LMArena Instruction Following | 1405 | 1437 |
Long Context Too close to call
GPT-5.4 mini: 43.0 (#112), GPT-5.6 Luna: 43.9 (#82)
| Benchmark | GPT-5.4 mini | GPT-5.6 Luna |
|---|---|---|
| LMArena Longer Query | 1407 | 1436 |
Writing & Preference GPT-5.6 Luna leads
GPT-5.4 mini: 64.0 (#58), GPT-5.6 Luna: 68.0 (#29)
| Benchmark | GPT-5.4 mini | GPT-5.6 Luna |
|---|---|---|
| LMArena Text | 1412 | 1431 |
| LMArena Creative Writing | 1370 | 1396 |
| EQ-Bench Creative Writing | 1665 | 1829 |
| LMArena Multi-Turn | 1429 | 1434 |
| EQ-Bench 4 | — | 1156 |
Frequently asked questions
Is GPT-5.4 mini better than GPT-5.6 Luna?
GPT-5.6 Luna is the stronger model overall, scoring 54.6 to 45.0 on the Noometry Index.
Which is cheaper, GPT-5.4 mini or GPT-5.6 Luna?
GPT-5.6 Luna is cheaper. It lists at $0.20 per million input tokens and $1.20 per million output tokens; GPT-5.4 mini lists at $0.75 and $4.50.
Is GPT-5.4 mini or GPT-5.6 Luna better for coding?
GPT-5.6 Luna scores higher on coding benchmarks: 54.5 versus 45.2 in the Noometry coding category.
Which has the bigger context window?
GPT-5.6 Luna does, with 1.05M tokens against 400K.
How many benchmarks do GPT-5.4 mini and GPT-5.6 Luna share?
40 benchmarks have published results for both models. GPT-5.4 mini has 46 scored results on Noometry and GPT-5.6 Luna has 52.