OpenAI
availableShows if the model has enough results for an index.GPT-OSS 120B
GPT-OSS 120B is a non-reasoning model from OpenAI in the GPT-OSS family. 27 benchmarks count toward its score, in 7 categories.
IndexOverall score. 50 is the middle.44.3 ±4.1
CoverageShare of the index weight with results.90%
SpeedOutput tokens per second.44/s
Input / 1MUS dollars per 1M input tokens.$0.15
Output / 1MUS dollars per 1M output tokens.$0.6
ContextMaximum tokens in one request.131K
EloLMArena rating and rank.1366 (#176)
50 is the middle of the board. The range shows the doubt in the index.
29,955 votes. Elo shows what people prefer. It does not change the score.
CapabilitiesScore per category. 50 is the middle.
50 is the middleResults
27 counted| BenchmarkThe test name. | CategoryThe capability that the test measures. | ResultThe score from the publisher. | IndexThis result on the index scale. | RunThe settings of the run. | DateDate of the result. | Published byThe source of the result. |
|---|---|---|---|---|---|---|
| MATH 500 | Math | 94.8% | 47.7 | — | 9 Jan 2026 | Vals AI |
| AIME | Math | 92.6% | 56.9 | — | 16 Apr 2026 | Vals AI |
| MGSM | Multilingual | 92.0% | — | — | 9 Jan 2026 | Vals AI |
| OTIS Mock AIME 2024-2025 | Math | 88.9% | 61.2 | high effort | — | Epoch AI |
| LiveCodeBench | Coding | 83.2% | 60.0 | — | 1 Sept 2026 | Vals AI |
| MMLU Pro | Knowledge | 79.2% | 45.2 | — | 1 Sept 2026 | Vals AI |
| GPQA Diamond | Knowledge | 78.5% | 50.7 | — | 1 Sept 2026 | Vals AI |
| Artificial Analysis GPQA Diamond | Knowledge | 78.2% | 48.7 | — | — | Artificial Analysis |
| GPQA diamond | Knowledge | 75.8% | 48.2 | high effort | — | Epoch AI |
| React Native Evals | Coding | 71.6% | 47.9 | — | — | Callstack |
| Artificial Analysis IFBench | Instruction | 69.0% | 59.1 | — | — | Artificial Analysis |
| τ²-Bench Tool-Agent-User Evaluation | Agentic | 65.8% | 46.0 | — | — | Victor Barres et al. |
| EuroEval Swedish | Multilingual | 61.3% | 78.8 | — | — | EuroEval |
| EuroEval Portuguese | Multilingual | 59.3% | 76.3 | — | — | EuroEval |
| EuroEval Italian | Multilingual | 58.4% | 75.2 | — | — | EuroEval |
| EuroEval French | Multilingual | 58.4% | 75.2 | — | — | EuroEval |
| EuroEval Dutch | Multilingual | 57.8% | 74.5 | — | — | EuroEval |
| EuroEval Polish | Multilingual | 57.6% | 74.2 | — | — | EuroEval |
| EuroEval Spanish | Multilingual | 53.5% | 69.1 | — | — | EuroEval |
| Artificial Analysis Long Context Reasoning | Reasoning | 52.0% | 44.2 | — | — | Artificial Analysis |
| EuroEval German | Multilingual | 51.2% | 66.2 | — | — | EuroEval |
| Artificial Analysis SciCode | Coding | 34.0% | 40.1 | — | — | Artificial Analysis |
| SWE-bench | Coding | 33.6% | 24.5 | — | 1 Sept 2026 | Vals AI |
| Artificial Analysis Coding Index | Coding | 30.4% | 40.4 | — | — | Artificial Analysis |
| Gert Labs Composite Game Benchmark | Agentic | 29.6% | 38.9 | — | — | Gert Labs |
| SWE-bench Verified | Coding | 26.0% | 18.4 | mini-SWE-agent | 26 Feb 2026 | SWE-bench team |
| SWE-bench Verified | Coding | 26.0% | 18.4 | mini-SWE-agent | 1 Sept 2026 | SWE-bench team |
| Terminal-Bench 1.0 | Agentic | 22.5% | 30.6 | — | 12 Jan 2026 | Vals AI |
| Artificial Analysis Omniscience Accuracy | Knowledge | 21.8% | 40.8 | — | — | Artificial Analysis |
| Chess Puzzles | Reasoning | 20.0% | 49.4 | high effort | — | Epoch AI |
| Artificial Analysis Humanity's Last Exam | Knowledge | 19.6% | 45.8 | — | — | Artificial Analysis |
| Terminal-Bench 2.0 | Agentic | 19.1% | 36.3 | — | 4 Jun 2026 | Vals AI |
| Artificial Analysis Intelligence Index | Knowledge | 11.6% | 36.9 | — | — | Artificial Analysis |
| Artificial Analysis Agentic Index | Agentic | 6.2% | 42.2 | — | — | Artificial Analysis |
| GDPval-AA normalized | Agentic | 4.8% | 39.3 | — | — | Artificial Analysis |
| APEX-Agents-AA | Agentic | 3.1% | 43.5 | — | — | Artificial Analysis / Mercor |
| Critical Physics Tasks | Reasoning | 1.1% | 40.7 | — | — | Artificial Analysis |
| Mystery Game Puzzles | Reasoning | 0.0% | 33.9 | high effort | — | Epoch AI |
27 benchmarks count, from 37 of 38 results. A grey row does not count. Too few models took that benchmark.