GPT-OSS-120B: benchmark results
OpenAI's open-weight GPT-OSS 120B, an Apache-2.0 sparse MoE (~5B active) reasoning model (August 2025). Provider: OpenAI. Released 2025-08-05. Access: Open.
Aggregate Rating 1636 ± 3, rank #341 of 1656 rated models, from 2897 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AgentJudgeBench (Hard, With Ground Truth) - SmolLM3-3B Outputs | 77.8 | Judge alignment (%; agreement of the judge's per-metric verd | 100 |
| Bengali Dialect QA Consistency - Mymensingh | 9.22 | Dialect consistency score (0-10): a Gemini 2.5 Flash judge ( | 100 |
| BhashaBench - Finance | 71.05 | Accuracy (%) | 100 |
| BhashaBench - Finance - English | 74.11 | Accuracy (%) | 100 |
| BhashaBench - Finance - Hindi | 64.16 | Accuracy (%) | 100 |
| CelloAI-Doc-Bench - Expert Similarity | 60.2 | Mean embedding cosine similarity (-1 to 1, times 100) betwee | 100 |
| CelloAI-Doc-Bench - Tag Coverage F1 | 95.2 | Tag-level F1 (0-1, times 100) of the @param and @return tags | 100 |
| DTap-Bench (Direct Attacks) - Browser | 10 | Attack success rate (%; 85 direct red-teaming web browser ta | 100 |
| DTap-Bench (Indirect Attacks) - CRM | 9.3 | Attack success rate (%; 150 indirect red-teaming customer re | 100 |
| DTap-Bench (Indirect Attacks) - Customer Service | 21.4 | Attack success rate (%; 101 indirect red-teaming customer se | 100 |
| DTap-Bench (Indirect Attacks) - Workflow | 7.2 | Attack success rate (%; 107 indirect red-teaming workflow (e | 100 |
| ESF-Bench - Long and Multi-Value | 81.5 | Tagged slot accuracy (%; slots tagged with a LMV (Long and M | 100 |
Interactive version: theaggregate.ai/model?slug=gpt-oss-120b · How It Works · Data refreshed daily, snapshot 2026-10-20.