GPT-5.4 scores 47.2 on the Respan Index (#28), ahead of DeepSeek V4 Pro at 46.9 (#29). GPT-5.4 leads on 14 of the 24 benchmarks both report. DeepSeek V4 Pro is 2.8x cheaper per token at list price. GPT-5.4 has the larger context window (1.1M tokens).
* Estimated: no published scores in that category. How the Respan Index works
| 45.3% |
| 78.6% |
| GPQA Diamond | 90.9% | 92.8% |
| LiveBench Agentic Coding (LiveBench) | 54.9% | 53.8% |
| LiveBench Coding (LiveBench) | 77.2% | 77.5% |
| LiveBench Data Analysis (LiveBench) | 79.2% | 79.3% |
| LiveBench Instruction Following (LiveBench) | 67.7% | 70.2% |
| LiveBench Language (LiveBench) | 82.1% | 82.6% |
| LiveBench Mathematics (LiveBench) | 95.1% | 94.1% |
| LiveBench Reasoning (LiveBench) | 85.8% | 88.1% |
| Mystery Game Puzzles (Epoch AI) | 43% | 37% |
| OTIS Mock AIME 2024-2025 (Epoch AI) | 96.7% | 97.8% |
| SimpleQA Verified | 47% | 45.1% |
| SWE-bench Verified (Epoch AI) | 77.6% | 76.9% |
| Vending-Bench 2 (Andon Labs) | 3284.52 | 6144.18 |
| WeirdML (Håvard Tveit Ihle) | 66.2% | 77.7% |
| LMArena Agent (LMArena) | 0.0119 | 0.0105 |
| LMArena Elo | 1451 | 1453 |
| LMArena WebDev (LMArena) | 1583 | 1395 |
| AIME 2026 | 96.7% | 99.2% |
| Humanity's Last Exam (with tools) | 60% | 52.1% |
Higher is better. Each score links to where it was published.