# GPT-5.2 vs GPT-5.4

> Side-by-side benchmark comparison for GPT-5.2 and GPT-5.4 across agentic, coding, multimodal, knowledge, reasoning, multilingual, and math tasks.

- Canonical page: https://benchlm.ai/compare/gpt-5-2-vs-gpt-5-4
- Last updated: October 17, 2026

- Shared sourced benchmarks: 15
- HTML indexing: indexable
- Ranking lane: BenchAlign v5.8

## Quick Verdict

Pick GPT-5.4 if you want the stronger benchmark profile. GPT-5.2 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 66.7 to 59.86.
- The clearest category separation is in coding, where the averages are 36.8 for GPT-5.2 and 47.7 for GPT-5.4.
- The biggest single benchmark swing is OSWorld-Verified in Agentic, with scores of 47.3% and 75%.
- GPT-5.2 is the cheaper option on output tokens, which matters if you expect large responses or heavy interactive use.
- GPT-5.4 also has the larger context window at 1.05M.

## Model Snapshot

| Property | GPT-5.2 | GPT-5.4 |
|----------|----------|----------|
| Creator | OpenAI | OpenAI |
| Type | Proprietary | Proprietary |
| Reasoning | Reasoning | Reasoning |
| Context | 400K | 1.05M |
| Overall Score | 59.86 | 66.7 |
| Benchmarks Covered | 31 | 42 |
| Pricing (input/output) | $1.75 / $14.00 | $2.50 / $15.00 |

## Category Breakdown

### Agentic

- Winner: GPT-5.4
- GPT-5.2 public-lane score: 44.4 (Supported · #58/128)
- GPT-5.4 public-lane score: 52.1 (Supported · #44/128)

| Benchmark | GPT-5.2 | GPT-5.4 | Winner |
|-----------|-----------|-----------|--------|
| BrowseComp | 65.8% | 82.7% | GPT-5.4 |
| OSWorld-Verified | 47.3% | 75% | GPT-5.4 |
| Gert Labs | 46.54% | 64.89% | GPT-5.4 |
| JobBench | 34.3% | 38.9% | GPT-5.4 |
| Terminal-Bench 2.0 | Coming soon | 75.1% | Coming soon |
| CyberGym | Coming soon | 79.0% | Coming soon |
| MCP Atlas | Coming soon | 70.6% | Coming soon |
| Toolathlon | Coming soon | 54.6% | Coming soon |
| τ²-bench results | Coming soon | 98.9% | Coming soon |
| Claw-Eval | Coming soon | 60.3% | Coming soon |
| DeepSearchQA | Coming soon | 73.6% | Coming soon |
| ResearchClawBench | Coming soon | 15.3% | Coming soon |
| ExploitGym | Coming soon | 6.0% | Coming soon |
| ApprenticeBench | Coming soon | 11% | Coming soon |
| MazeBench | Coming soon | 0% | Coming soon |

### Coding

- Winner: GPT-5.4
- GPT-5.2 public-lane score: 36.8 (Supported · #73/150)
- GPT-5.4 public-lane score: 47.7 (Supported · #48/150)

| Benchmark | GPT-5.2 | GPT-5.4 | Winner |
|-----------|-----------|-----------|--------|
| SWE-bench Verified | 80% | Coming soon | Coming soon |
| SWE-bench Pro | 55.6% | 57.7% | GPT-5.4 |
| Vibe Code Bench | 53.50% | 67.42% | GPT-5.4 |
| LiveCodeBench Pro | Coming soon | 87.5% | Coming soon |
| React Native Evals | Coming soon | 85.3% | Coming soon |
| Terminal-Bench 2.0 | Coming soon | 75.1% | Coming soon |
| PostTrainBench v1.1 | Coming soon | 19.0% | Coming soon |

### Reasoning

- Winner: Not comparable
- GPT-5.2 public-lane score: 64.3 (Unranked · 4 rankable rows)
- GPT-5.4 public-lane score: 64.2 (#23/28)

| Benchmark | GPT-5.2 | GPT-5.4 | Winner |
|-----------|-----------|-----------|--------|
| ARC-AGI-2 | 52.9% | 74.0% | GPT-5.4 |
| ARC-AGI-1 | 86.17% | 93.67% | GPT-5.4 |
| ARC-AGI-3 | Coming soon | 0.2% | Coming soon |

### Multimodal & Grounded

- Winner: Directional only
- GPT-5.2 public-lane score: 75.9 (#22/54)
- GPT-5.4 public-lane score: 73.1 (#30/54)

| Benchmark | GPT-5.2 | GPT-5.4 | Winner |
|-----------|-----------|-----------|--------|
| MMMU | 86.7% | Coming soon | Coming soon |
| MMMU-Pro | 79.5% | 81.2% | GPT-5.4 |
| RealWorldQA | 83.3% | Coming soon | Coming soon |
| OmniDocBench 1.5 | 85.7% | Coming soon | Coming soon |
| Video-MME (with subtitle) | 86.0% | Coming soon | Coming soon |
| Video-MME (w/o subtitle) | 85.8% | Coming soon | Coming soon |
| MathVision | 83.0% | Coming soon | Coming soon |
| We-Math | 79.0% | Coming soon | Coming soon |
| DynaMath | 86.8% | Coming soon | Coming soon |
| SimpleVQA | 55.8% | 61.1% | GPT-5.4 |
| CC-OCR | 70.3% | Coming soon | Coming soon |
| AI2D_TEST | 92.2% | Coming soon | Coming soon |
| CountBench | 91.9% | Coming soon | Coming soon |
| ERQA | 59.8% | 65.4% | GPT-5.4 |
| VideoMMMU | 85.9% | Coming soon | Coming soon |
| MLVU (M-Avg) | 85.6% | Coming soon | Coming soon |
| CharXiv | 82.1% | 82.8% | GPT-5.4 |
| V* | 75.9% | Coming soon | Coming soon |
| OfficeQA Pro | Coming soon | 53.2% | Coming soon |
| MMMU-Pro w/ Python | Coming soon | 82.1% | Coming soon |
| ScreenSpot Pro | Coming soon | 85.4% | Coming soon |
| ZeroBench | Coming soon | 41.0% | Coming soon |
| MedXpertQA (MM) | Coming soon | 77.1% | Coming soon |

### Knowledge

- Winner: GPT-5.4
- GPT-5.2 public-lane score: 59.7 (Supported · #50/184)
- GPT-5.4 public-lane score: 69.1 (Supported · #20/184)

| Benchmark | GPT-5.2 | GPT-5.4 | Winner |
|-----------|-----------|-----------|--------|
| GPQA | 92.4% | 92.8% | GPT-5.4 |
| HLE | Coming soon | 52.1% | Coming soon |
| HLE w/o tools | Coming soon | 39.8% | Coming soon |
| GPQA-D | Coming soon | 92.8% | Coming soon |
| HealthBench Hard | Coming soon | 40.1% | Coming soon |
| MedXpertQA (Text) | Coming soon | 59.6% | Coming soon |
| HealthBench Professional | Coming soon | 48.1% | Coming soon |

### Instruction Following

- Winner: Directional only
- GPT-5.2 public-lane score: 92.6 (#19/148)
- GPT-5.4 public-lane score: 90.7 (#23/148)

### Mathematics

- Winner: Not comparable
- GPT-5.2 public-lane score: 71.1 (Unranked · 3 rankable rows)
- GPT-5.4 public-lane score: 82 (Unranked · 2 rankable rows)

| Benchmark | GPT-5.2 | GPT-5.4 | Winner |
|-----------|-----------|-----------|--------|
| AIME 2025 | 100% | Coming soon | Coming soon |
| FrontierMath v2 (Tiers 1-3) | 67.400% | 78.596% | GPT-5.4 |
| FrontierMath v2 (Tier 4) | 31.700% | 49.000% | GPT-5.4 |

## FAQ

### Which is better overall, GPT-5.2 or GPT-5.4?

GPT-5.4 is ahead overall on BenchLM right now.

### Where is the biggest gap between GPT-5.2 and GPT-5.4?

The widest category gap is in coding, where the averages are 36.8 for GPT-5.2 and 47.7 for GPT-5.4.

### How many benchmarks does GPT-5.2 cover on BenchLM?

GPT-5.2 currently has 31 sourced benchmark scores on BenchLM.

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

GPT-5.4 currently has 42 sourced benchmark scores on BenchLM.

## Related Comparisons

- [GPT-5.2 vs GPT-5.2 Pro](/compare/gpt-5-2-vs-gpt-5-2-pro)
- [GPT-5.4 vs GPT-5.2 Pro](/compare/gpt-5-2-pro-vs-gpt-5-4)
- [GPT-5.2 vs GPT-5.2 Instant](/compare/gpt-5-2-vs-gpt-5-2-instant)
- [GPT-5.4 vs GPT-5.2 Instant](/compare/gpt-5-2-instant-vs-gpt-5-4)

## Explore More

- [GPT-5.2 profile](/models/gpt-5-2)
- [GPT-5.4 profile](/models/gpt-5-4)
- [Compare Pricing](/llm-pricing)
- [Alternative Finder](/tools/alternative-finder)
- [LLM Selector](/tools/llm-selector)
- [Overall Rankings](/best/overall)