# GPT-5.4 vs Hy3 Preview

> Side-by-side benchmark comparison for GPT-5.4 and Hy3 Preview across agentic, coding, multimodal, knowledge, reasoning, multilingual, and math tasks.

- Canonical page: https://benchlm.ai/compare/gpt-5-4-vs-hy3-preview
- Last updated: September 23, 2026

- Shared sourced benchmarks: 4
- HTML indexing: indexable
- Ranking lane: BenchAlign v5.6

## Quick Verdict

Pick GPT-5.4 if you want the stronger benchmark profile. Hy3 Preview 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 69.23 to 45.32.
- The clearest category separation is in instruction following, where the averages are 89.2 for GPT-5.4 and 46.9 for Hy3 Preview.
- The biggest single benchmark swing is Gert Labs in Agentic, with scores of 64.89% and 36.91%.
- Hy3 Preview 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-5.4 | Hy3 Preview |
|----------|----------|----------|
| Creator | OpenAI | Tencent |
| Type | Proprietary | Open Weight |
| Reasoning | Reasoning | Reasoning |
| Context | 1.05M | 256K |
| Overall Score | 69.23 | 45.32 |
| Benchmarks Covered | 38 | 6 |
| Pricing (input/output) | $2.50 / $15.00 | $0.00 / $0.00 |

## Category Breakdown

### Agentic

- Winner: Not comparable
- GPT-5.4 public-lane score: 49.5 (Supported · #36/105)
- Hy3 Preview public-lane score: Coming soon

| Benchmark | GPT-5.4 | Hy3 Preview | Winner |
|-----------|-----------|-----------|--------|
| Terminal-Bench 2.0 | 75.1% | 54.4% | GPT-5.4 |
| CyberGym | 79.0% | Coming soon | Coming soon |
| BrowseComp | 82.7% | Coming soon | Coming soon |
| OSWorld-Verified | 75% | Coming soon | Coming soon |
| MCP Atlas | 70.6% | Coming soon | Coming soon |
| Toolathlon | 54.6% | Coming soon | Coming soon |
| 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% | 36.91% | GPT-5.4 |
| ResearchClawBench | 15.3% | Coming soon | Coming soon |
| JobBench | 38.9% | Coming soon | Coming soon |
| ExploitGym | 6.0% | Coming soon | Coming soon |
| ApprenticeBench | 11% | Coming soon | Coming soon |

### Coding

- Winner: Directional only
- GPT-5.4 public-lane score: 51 (Supported · #38/135)
- Hy3 Preview public-lane score: 33.3 (Estimated · #85/135)

| Benchmark | GPT-5.4 | Hy3 Preview | Winner |
|-----------|-----------|-----------|--------|
| LiveCodeBench Pro | 87.5% | Coming soon | Coming soon |
| SWE-bench Pro | 57.7% | Coming soon | Coming soon |
| React Native Evals | 85.3% | Coming soon | Coming soon |
| Vibe Code Bench | 67.42% | Coming soon | Coming soon |
| SWE-bench Verified | Coming soon | 74.4% | Coming soon |
| Terminal-Bench 2.0 | Coming soon | 54.4% | Coming soon |

### Reasoning

- Winner: Not comparable
- GPT-5.4 public-lane score: 57.2 (#15/18)
- Hy3 Preview public-lane score: Coming soon

| Benchmark | GPT-5.4 | Hy3 Preview | Winner |
|-----------|-----------|-----------|--------|
| ARC-AGI-2 | 74.0% | Coming soon | Coming soon |
| ARC-AGI-3 | 0.2% | Coming soon | Coming soon |

### Multimodal & Grounded

- Winner: Not comparable
- GPT-5.4 public-lane score: 69.3 (#21/50)
- Hy3 Preview public-lane score: Coming soon

| Benchmark | GPT-5.4 | Hy3 Preview | Winner |
|-----------|-----------|-----------|--------|
| MMMU-Pro | 81.2% | Coming soon | Coming soon |
| OfficeQA Pro | 53.2% | Coming soon | Coming soon |
| MMMU-Pro w/ Python | 82.1% | Coming soon | Coming soon |
| 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: Directional only
- GPT-5.4 public-lane score: 66.5 (Supported · #17/160)
- Hy3 Preview public-lane score: 39.1 (Estimated · #93/160)

| Benchmark | GPT-5.4 | Hy3 Preview | Winner |
|-----------|-----------|-----------|--------|
| GPQA | 92.8% | 87.2% | GPT-5.4 |
| HLE | 52.1% | Coming soon | Coming soon |
| HLE w/o tools | 39.8% | Coming soon | Coming soon |
| GPQA-D | 92.8% | 87.2% | GPT-5.4 |
| HealthBench Hard | 40.1% | Coming soon | Coming soon |
| MedXpertQA (Text) | 59.6% | Coming soon | Coming soon |
| HealthBench Professional | 48.1% | Coming soon | Coming soon |

### Instruction Following

- Winner: Directional only
- GPT-5.4 public-lane score: 89.2 (#19/124)
- Hy3 Preview public-lane score: 46.9 (#86/124)

### Mathematics

- Winner: Not comparable
- GPT-5.4 public-lane score: 64.4 (Unranked · 2 rankable rows)
- Hy3 Preview public-lane score: Coming soon

| Benchmark | GPT-5.4 | Hy3 Preview | Winner |
|-----------|-----------|-----------|--------|
| FrontierMath v2 (Tiers 1-3) | 47.600% | Coming soon | Coming soon |
| FrontierMath v2 (Tier 4) | 27.100% | Coming soon | Coming soon |

## FAQ

### Which is better overall, GPT-5.4 or Hy3 Preview?

GPT-5.4 is ahead overall on BenchLM right now.

### Where is the biggest gap between GPT-5.4 and Hy3 Preview?

The widest category gap is in instruction following, where the averages are 89.2 for GPT-5.4 and 46.9 for Hy3 Preview.

### 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 Hy3 Preview cover on BenchLM?

Hy3 Preview currently has 6 sourced benchmark scores on BenchLM.

## Related Comparisons

- [GPT-5.4 vs GPT-5.4 Pro](/compare/gpt-5-4-vs-gpt-5-4-pro)
- [Hy3 Preview vs GPT-5.4 Pro](/compare/gpt-5-4-pro-vs-hy3-preview)
- [GPT-5.4 vs GPT-5.4 mini](/compare/gpt-5-4-vs-gpt-5-4-mini)
- [Hy3 Preview vs GPT-5.4 mini](/compare/gpt-5-4-mini-vs-hy3-preview)

## Explore More

- [GPT-5.4 profile](/models/gpt-5-4)
- [Hy3 Preview profile](/models/hy3-preview)
- [Compare Pricing](/llm-pricing)
- [Alternative Finder](/tools/alternative-finder)
- [LLM Selector](/tools/llm-selector)
- [Overall Rankings](/best/overall)