Compare / Gemma 4 31B vs gpt-oss-20b
Gemma 4 31B vs gpt-oss-20b
21 results from 3 sources that measured both models. Each row is in the source's own units; there is no overall winner.
| Benchmark | Metric | Gemma 4 31B | gpt-oss-20b |
|---|---|---|---|
| Arena (formerly LMArena)reported | Business, management and financeArena rating | 1,449 | 1,323 |
| Arena (formerly LMArena)reported | Creative writingArena rating | 1,419 | 1,240 |
| Arena (formerly LMArena)reported | Expert promptsArena rating | 1,499 | 1,314 |
| Arena (formerly LMArena)reported | Instruction followingArena rating | 1,453 | 1,281 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,453 | 1,318 |
| Arena (formerly LMArena)reported | Writing, literature and languageArena rating | 1,435 | 1,275 |
| Artificial Analysisreported | Artificial Analysis Coding Indexindex score | 43.4 | 20.7 |
| Artificial Analysisreported | Artificial Analysis Intelligence Indexindex score | 14.7 | 10.0 |
| Artificial Analysisreported | GPQA Diamond% of questions | 85.7% | 68.8% |
| Artificial Analysisreported | Humanity's Last Exam% of questions | 23.6% | 11.0% |
| Artificial Analysisreported | IFBench% of instructions | 75.6% | 65.1% |
| Artificial Analysisreported | Long-context reasoning (AA-LCR)% of questions | 69.7% | 34.7% |
| Artificial Analysisreported | SciCode% of problems | 45.5% | 38.9% |
| Artificial Analysisreported | Terminal-Bench 2.1% of tasks | 43.4% | 13.9% |
| Artificial Analysisreported | Terminal-Bench 4.0% of tasks | 0.0% | 0.0% |
| Artificial Analysisreported | Terminal-Bench Hard% of tasks | 36.4% | 10.6% |
| Artificial Analysisreported | Τ-bench banking% of tasks | 14.8% | 7.0% |
| Artificial Analysisreported | Τ²-bench telecom% of tasks | 65.5% | 60.2% |
| UGI Leaderboardreported | Requested-length error% off the requested word count | 5.0% | 98.0% |
| UGI Leaderboardreported | Style adherencescore from 0 to 1 | 0.36 | 0.31 |
| UGI Leaderboardreported | Writing scorescore out of 100 | 43.1 | 24.6 |
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.