# GPT-5.4 vs GPT-5.6 Luna

> Side-by-side benchmark comparison for GPT-5.4 and GPT-5.6 Luna across agentic, coding, multimodal, knowledge, reasoning, multilingual, and math tasks.

- Canonical page: https://benchlm.ai/compare/gpt-5-4-vs-gpt-5-6-luna
- Last updated: September 24, 2026

- Shared sourced benchmarks: 16
- HTML indexing: indexable
- Ranking lane: BenchAlign v5.7

## Quick Verdict

Pick GPT-5.4 if you want the stronger benchmark profile. GPT-5.6 Luna only makes more sense when its price, context window, or workload-specific category wins matter more than the overall score.

## Summary

- GPT-5.4 leads overall 68.51 to 65.6.
- The clearest category separation is in mathematics, where the averages are 64.4 for GPT-5.4 and 94.2 for GPT-5.6 Luna.
- The biggest single benchmark swing is FrontierMath v2 (Tier 4) in Mathematics, with scores of 27.100% and 58.500%.
- GPT-5.6 Luna is the cheaper option on output tokens, which matters if you expect large responses or heavy interactive use.

## Model Snapshot

| Property | GPT-5.4 | GPT-5.6 Luna |
|----------|----------|----------|
| Creator | OpenAI | OpenAI |
| Type | Proprietary | Proprietary |
| Reasoning | Reasoning | Reasoning |
| Context | 1.05M | 1.05M |
| Overall Score | 68.51 | 65.6 |
| Benchmarks Covered | 38 | 29 |
| Pricing (input/output) | $2.50 / $15.00 | $0.20 / $1.20 |

## Category Breakdown

### Agentic

- Winner: GPT-5.6 Luna
- GPT-5.4 public-lane score: 49.4 (Supported · #36/105)
- GPT-5.6 Luna public-lane score: 55.3 (Supported · #28/105)

| Benchmark | GPT-5.4 | GPT-5.6 Luna | Winner |
|-----------|-----------|-----------|--------|
| Terminal-Bench 2.0 | 75.1% | Coming soon | Coming soon |
| CyberGym | 79.0% | 77.9% | GPT-5.4 |
| BrowseComp | 82.7% | 83.3% | GPT-5.6 Luna |
| OSWorld-Verified | 75% | Coming soon | Coming soon |
| MCP Atlas | 70.6% | Coming soon | Coming soon |
| Toolathlon | 54.6% | 53.4% | GPT-5.4 |
| τ²-bench results | 98.9% | Coming soon | Coming soon |
| Claw-Eval | 60.3% | Coming soon | Coming soon |
| DeepSearchQA | 73.6% | Coming soon | Coming soon |
| Gert Labs | 64.89% | Coming soon | Coming soon |
| ResearchClawBench | 15.3% | Coming soon | Coming soon |
| JobBench | 38.9% | Coming soon | Coming soon |
| ExploitGym | 6.0% | 12.4% | GPT-5.6 Luna |
| ApprenticeBench | 11% | 7% | GPT-5.4 |
| Terminal-Bench 3.0 | Coming soon | 14.3% | Coming soon |
| Terminal-Bench 2.1 | Coming soon | 84.7% | Coming soon |
| OSWorld 2.0 | Coming soon | 45.6% | Coming soon |
| Terminal-Bench 2.1 (Vals) | Coming soon | 79.0% | Coming soon |

### Coding

- Winner: GPT-5.6 Luna
- GPT-5.4 public-lane score: 50.7 (Supported · #38/135)
- GPT-5.6 Luna public-lane score: 64.5 (Supported · #9/135)

| Benchmark | GPT-5.4 | GPT-5.6 Luna | Winner |
|-----------|-----------|-----------|--------|
| LiveCodeBench Pro | 87.5% | Coming soon | Coming soon |
| SWE-bench Pro | 57.7% | 62.7% | GPT-5.6 Luna |
| React Native Evals | 85.3% | Coming soon | Coming soon |
| Vibe Code Bench | 67.42% | Coming soon | Coming soon |
| Terminal-Bench 2.1 | Coming soon | 84.7% | Coming soon |
| DeepSWE | Coming soon | 67.2% | Coming soon |
| FrontierCode 1.1 Extended | Coming soon | 55.1% | Coming soon |
| cursorBench32 | Coming soon | 61.1% | Coming soon |
| VulcanBench v3 | Coming soon | 85.5% | Coming soon |
| SWE-bench (Vals) | Coming soon | 93.0% | Coming soon |

### Reasoning

- Winner: GPT-5.4
- GPT-5.4 public-lane score: 60.5 (#14/19)
- GPT-5.6 Luna public-lane score: 54.7 (#18/19)

| Benchmark | GPT-5.4 | GPT-5.6 Luna | Winner |
|-----------|-----------|-----------|--------|
| ARC-AGI-2 | 74.0% | 59.5% | GPT-5.4 |
| ARC-AGI-3 | 0.2% | 0.2% | GPT-5.4 |

### Multimodal & Grounded

- Winner: Directional only
- GPT-5.4 public-lane score: 69.3 (#21/50)
- GPT-5.6 Luna public-lane score: 67.1 (#22/50)

| Benchmark | GPT-5.4 | GPT-5.6 Luna | Winner |
|-----------|-----------|-----------|--------|
| MMMU-Pro | 81.2% | 78.4% | GPT-5.4 |
| OfficeQA Pro | 53.2% | Coming soon | Coming soon |
| MMMU-Pro w/ Python | 82.1% | 79.5% | GPT-5.4 |
| CharXiv | 82.8% | Coming soon | Coming soon |
| ERQA | 65.4% | Coming soon | Coming soon |
| SimpleVQA | 61.1% | Coming soon | Coming soon |
| ScreenSpot Pro | 85.4% | Coming soon | Coming soon |
| ZeroBench | 41.0% | Coming soon | Coming soon |
| MedXpertQA (MM) | 77.1% | Coming soon | Coming soon |

### Knowledge

- Winner: GPT-5.4
- GPT-5.4 public-lane score: 66.8 (Supported · #16/158)
- GPT-5.6 Luna public-lane score: 64.6 (Supported · #22/158)

| Benchmark | GPT-5.4 | GPT-5.6 Luna | Winner |
|-----------|-----------|-----------|--------|
| GPQA | 92.8% | 92.3% | GPT-5.4 |
| HLE | 52.1% | Coming soon | Coming soon |
| HLE w/o tools | 39.8% | Coming soon | Coming soon |
| GPQA-D | 92.8% | 92.3% | GPT-5.4 |
| HealthBench Hard | 40.1% | 32.0% | GPT-5.4 |
| MedXpertQA (Text) | 59.6% | Coming soon | Coming soon |
| HealthBench Professional | 48.1% | 55.7% | GPT-5.6 Luna |
| GPQA Diamond (Vals) | Coming soon | 91.7% | Coming soon |
| MMLU-Pro (Vals) | Coming soon | 86.0% | Coming soon |

### Instruction Following

- Winner: Not comparable
- GPT-5.4 public-lane score: 89.2 (#19/124)
- GPT-5.6 Luna public-lane score: Coming soon

### Mathematics

- Winner: Not comparable
- GPT-5.4 public-lane score: 64.4 (Unranked · 2 rankable rows)
- GPT-5.6 Luna public-lane score: 94.2 (Unranked · 3 rankable rows)

| Benchmark | GPT-5.4 | GPT-5.6 Luna | Winner |
|-----------|-----------|-----------|--------|
| FrontierMath v2 (Tiers 1-3) | 47.600% | 78.600% | GPT-5.6 Luna |
| FrontierMath v2 (Tier 4) | 27.100% | 58.500% | GPT-5.6 Luna |
| FrontierMath (legacy) | Coming soon | 78.6% | Coming soon |

## FAQ

### Which is better overall, GPT-5.4 or GPT-5.6 Luna?

GPT-5.4 is ahead overall on BenchLM right now.

### Where is the biggest gap between GPT-5.4 and GPT-5.6 Luna?

The widest category gap is in mathematics, where the averages are 64.4 for GPT-5.4 and 94.2 for GPT-5.6 Luna.

### How many benchmarks does GPT-5.4 cover on BenchLM?

GPT-5.4 currently has 38 sourced benchmark scores on BenchLM.

### How many benchmarks does GPT-5.6 Luna cover on BenchLM?

GPT-5.6 Luna currently has 29 sourced benchmark scores on BenchLM.

## Related Comparisons

- [GPT-5.4 vs GPT-5.4 Pro](/compare/gpt-5-4-vs-gpt-5-4-pro)
- [GPT-5.6 Luna vs GPT-5.4 Pro](/compare/gpt-5-4-pro-vs-gpt-5-6-luna)
- [GPT-5.4 vs GPT-5.4 mini](/compare/gpt-5-4-vs-gpt-5-4-mini)
- [GPT-5.6 Luna vs GPT-5.4 mini](/compare/gpt-5-4-mini-vs-gpt-5-6-luna)

## Explore More

- [GPT-5.4 profile](/models/gpt-5-4)
- [GPT-5.6 Luna profile](/models/gpt-5-6-luna)
- [Compare Pricing](/llm-pricing)
- [Alternative Finder](/tools/alternative-finder)
- [LLM Selector](/tools/llm-selector)
- [Overall Rankings](/best/overall)