Compare / GPT-5.4 Mini vs GPT-5.6 Terra
GPT-5.4 Mini vs GPT-5.6 Terra
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 | GPT-5.4 Mini | GPT-5.6 Terra |
|---|---|---|---|
| Arena (formerly LMArena)reported | OverallArena rating | 1,447 | 1,467 |
| Arena (formerly LMArena)reported | Business, management and financeArena rating | 1,460 | 1,466 |
| Arena (formerly LMArena)reported | Creative writingArena rating | 1,402 | 1,428 |
| Arena (formerly LMArena)reported | Expert promptsArena rating | 1,480 | 1,504 |
| Arena (formerly LMArena)reported | Instruction followingArena rating | 1,434 | 1,462 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,447 | 1,465 |
| Arena (formerly LMArena)reported | Writing, literature and languageArena rating | 1,424 | 1,445 |
| Artificial Analysisreported | Artificial Analysis Coding Indexindex score | 56.1 | 76.7 |
| Artificial Analysisreported | Artificial Analysis Intelligence Indexindex score | 24.1 | 42.1 |
| Artificial Analysisreported | GPQA Diamond% of questions | 87.5% | 92.5% |
| Artificial Analysisreported | Humanity's Last Exam% of questions | 28.1% | 42.9% |
| Artificial Analysisreported | IFBench% of instructions | 73.3% | 71.2% |
| Artificial Analysisreported | Long-context reasoning (AA-LCR)% of questions | 77.0% | 83.0% |
| Artificial Analysisreported | SciCode% of problems | 52.1% | 55.0% |
| Artificial Analysisreported | Terminal-Bench 2.1% of tasks | 59.2% | 88.0% |
| Artificial Analysisreported | Terminal-Bench 4.0% of tasks | 2.0% | 35.4% |
| Artificial Analysisreported | Terminal-Bench Hard% of tasks | 52.3% | 62.9% |
| Artificial Analysisreported | Τ-bench banking% of tasks | 25.6% | 40.2% |
| Artificial Analysisreported | Τ²-bench telecom% of tasks | 83.3% | 86.3% |
| SimpleQA Verified (Epoch AI)reported | Correct answers% of questions | 29.4% | 43.2% |
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.