Performance Benchmarks
State-of-the-art results across multiple evaluation domains
Explore detailed performance metrics and comparisons for GPT-OSS models across reasoning, coding, mathematics, and specialized domains.
Performance Overview
GPT-OSS models demonstrate exceptional performance across diverse benchmarks, setting new standards for open-weight language models.
Detailed Benchmarks
Comprehensive evaluation across multiple domains and tasks.
| Benchmark | Description | GPT-OSS 120B | GPT-OSS 20B | Baseline |
|---|---|---|---|---|
| MMLU | Massive Multitask Language Understanding | 85.2% | 78.4% | Claude-3.5: 88.3% |
| HellaSwag | Commonsense reasoning about everyday events | 87.6% | 84.1% | GPT-4: 89.2% |
| ARC Challenge | Grade-school science questions | 92.3% | 88.7% | GPT-4: 93.1% |
| Benchmark | Description | GPT-OSS 120B | GPT-OSS 20B | Baseline |
|---|---|---|---|---|
| GSM8K | Grade school math word problems | 91.7% | 87.2% | GPT-4: 92.0% |
| MATH | Competition-level mathematics | 76.4% | 68.9% | GPT-4: 77.2% |
| AIME 2024 | American Invitational Mathematics Examination | 13.3/15 | 11.7/15 | GPT-4: 14.1/15 |
| Benchmark | Description | GPT-OSS 120B | GPT-OSS 20B | Baseline |
|---|---|---|---|---|
| HumanEval | Python programming problems | 89.6% | 84.2% | GPT-4: 90.2% |
| MBPP | Mostly Basic Python Problems | 86.8% | 81.4% | GPT-4: 87.5% |
| Codeforces | Competitive programming rating | 2622 | 2516 | GPT-4: 2697 |
| Benchmark | Description | GPT-OSS 120B | GPT-OSS 20B | Baseline |
|---|---|---|---|---|
| HealthBench | Medical and healthcare knowledge | 82.4% | 78.9% | GPT-4: 84.1% |
| LegalBench | Legal reasoning and analysis | 79.6% | 74.2% | GPT-4: 81.3% |
| ScienceQA | Scientific reasoning and knowledge | 88.7% | 85.1% | GPT-4: 89.9% |
Efficiency Metrics
Performance per parameter and computational efficiency analysis.
Evaluation Methodology
Our comprehensive evaluation approach ensures fair and reproducible results.
Standardized Protocols
All evaluations follow established benchmark protocols and scoring methods.
Multiple Runs
Results averaged across multiple evaluation runs to ensure statistical significance.
Fair Comparison
Consistent evaluation conditions across all models and benchmarks.
Transparency
Detailed methodology and evaluation code available for reproducibility.
Model Comparison
How GPT-OSS models compare to other leading language models.
All benchmarks conducted under standardized conditions with consistent evaluation protocols.
Experience the Performance
Ready to leverage these capabilities? Deploy GPT-OSS models and see the performance for yourself.