Compare / Claude Opus 4.8 vs GPT-5.4 Mini
Claude Opus 4.8 vs GPT-5.4 Mini
20 results from 3 sources that measured both models. Each row is in the source's own units; there is no overall winner.
| Benchmark | Metric | Claude Opus 4.8 | GPT-5.4 Mini |
|---|---|---|---|
| Arena (formerly LMArena)reported | OverallArena rating | 1,469 | 1,447 |
| Arena (formerly LMArena)reported | Business, management and financeArena rating | 1,490 | 1,460 |
| Arena (formerly LMArena)reported | Creative writingArena rating | 1,467 | 1,402 |
| Arena (formerly LMArena)reported | Expert promptsArena rating | 1,525 | 1,480 |
| Arena (formerly LMArena)reported | Instruction followingArena rating | 1,491 | 1,434 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,480 | 1,447 |
| Arena (formerly LMArena)reported | Writing, literature and languageArena rating | 1,474 | 1,424 |
| Artificial Analysisreported | Artificial Analysis Coding Indexindex score | 74.3 | 56.1 |
| Artificial Analysisreported | Artificial Analysis Intelligence Indexindex score | 41.8 | 24.1 |
| Artificial Analysisreported | GPQA Diamond% of questions | 92.0% | 87.5% |
| Artificial Analysisreported | Humanity's Last Exam% of questions | 48.7% | 28.1% |
| Artificial Analysisreported | IFBench% of instructions | 62.2% | 73.3% |
| Artificial Analysisreported | Long-context reasoning (AA-LCR)% of questions | 77.7% | 77.0% |
| Artificial Analysisreported | SciCode% of problems | 54.4% | 52.1% |
| Artificial Analysisreported | Terminal-Bench 2.1% of tasks | 84.6% | 59.2% |
| Artificial Analysisreported | Terminal-Bench 4.0% of tasks | 21.7% | 2.0% |
| Artificial Analysisreported | Terminal-Bench Hard% of tasks | 58.3% | 52.3% |
| Artificial Analysisreported | Τ-bench banking% of tasks | 34.2% | 25.6% |
| Artificial Analysisreported | Τ²-bench telecom% of tasks | 94.4% | 83.3% |
| SimpleQA Verified (Epoch AI)reported | Correct answers% of questions | 53.0% | 29.4% |
Bold green marks the better value on that metric. Where a source reports ranges that overlap, the difference may not be meaningful; see the benchmark page for ranges.