Updated daily at 05:00 UTC
- Released
- 5 Aug 2025
- Status
- Current
- Context
- 131.1K
- Max output
- 131.1K
- Input per 1M tokens
- —
- Output per 1M tokens
- —
- Open weights
- Yes
- License
- apache-2.0
- Parameters
- 21B, 3.6B active
- Knowledge cutoff
- Jun 2024
- Reasoning
- Yes
- Modalities
- text → text
Category scores
Out of 100, from independent and third-party results only, by method category-2026-09-29.1.
| Category | Score | Rank | Coverage |
|---|---|---|---|
| Knowledge & reasoning | 8.5 | 99 | 2 of 4 |
| Human preference | 38.4 | Not ranked | 1 of 5 |
Benchmark results
Every result we hold for this model, with who ran it and where it was published.
Knowledge & reasoning
| Benchmark | Metric | Value | Provenance | Setting | Evaluated | Source |
|---|---|---|---|---|---|---|
| Chess Puzzles | Puzzles solved | 4.0% | Independent | medium | 27 Aug 2026 | Epoch AI |
| Chess Puzzles | Puzzles solved | 2.0% | Independent | low | 27 Aug 2026 | Epoch AI |
| Chess Puzzles | Puzzles solved | 0.0% | Independent | high | 27 Aug 2026 | Epoch AI |
| DTBench (Decision Theory Benchmark) | Accuracy (capability questions) | 68.0% | Third party | high | — | https://conceptualreasoning.ai/dtbench via Epoch AI |
| Humanity's Last Exam (HLE) | Accuracy | 17.3% | Self-reported | — | — | GPT-OSS Model Card via Hugging Face Community Evals |
| LMCA (Language Model Conceptual Argumentation) | Correlation with expert ratings (Pearson r) | 0.1451 | Third party | high | — | https://conceptualreasoning.ai/lmca via Epoch AI |
| MMLU-Pro | Accuracy | 73.6% | Self-reported | — | — | EvalEval via Hugging Face Community Evals |
Scientific reasoning
| Benchmark | Metric | Value | Provenance | Setting | Evaluated | Source |
|---|---|---|---|---|---|---|
| GPQA Diamond | Accuracy | 60.8% | Independent | medium | 27 Aug 2026 | Epoch AI |
| GPQA Diamond | Accuracy | 53.2% | Independent | low | 27 Aug 2026 | Epoch AI |
| GPQA Diamond | Accuracy | 46.0% | Independent | high | 6 Aug 2026 | Epoch AI |
| GPQA Diamond | Accuracy | 74.2% | Self-reported | — | — | GPT-OSS Model Card via Hugging Face Community Evals |
Mathematics
| Benchmark | Metric | Value | Provenance | Setting | Evaluated | Source |
|---|---|---|---|---|---|---|
| OTIS Mock AIME 2024-2025 | Accuracy | 65.3% | Independent | medium | 27 Aug 2026 | Epoch AI |
| OTIS Mock AIME 2024-2025 | Accuracy | 50.8% | Independent | high | 27 Aug 2026 | Epoch AI |
| OTIS Mock AIME 2024-2025 | Accuracy | 40.3% | Independent | low | 27 Aug 2026 | Epoch AI |
Coding & software engineering
| Benchmark | Metric | Value | Provenance | Setting | Evaluated | Source |
|---|---|---|---|---|---|---|
| Terminal-Bench 2.0 | Accuracy (mean over trials) | 3.4% | Third party | — | — | Terminal-Bench v2 Leaderboard via Epoch AI |
| Terminal-Bench 2.0 | Accuracy (mean over trials) | 3.4% | Third party | unknown | — | https://www.tbench.ai/leaderboard/terminal-bench/2.0 via Epoch AI |
| Terminal-Bench 2.0 | Accuracy (mean over trials) | 3.1% | Third party | — | — | Terminal-Bench v2 Leaderboard via Epoch AI |
| Terminal-Bench 2.0 | Accuracy (mean over trials) | 3.1% | Third party | unknown | — | https://www.tbench.ai/leaderboard/terminal-bench/2.0 via Epoch AI |
| WeirdML (v2) | Average accuracy | 40.9% | Third party | high | — | https://htihle.github.io/weirdml.html via Epoch AI |
| WeirdML (v2) | Average accuracy | 36.8% | Third party | medium | — | https://htihle.github.io/weirdml.html via Epoch AI |
Human preference
| Benchmark | Metric | Value | Provenance | Setting | Evaluated | Source |
|---|---|---|---|---|---|---|
| Text Arena (arena.ai) | Arena score (Elo scale) | 1318 | Independent | — | 25 Sep 2026 | Arena |
| Text Arena (arena.ai) | Votes | 10372 | Independent | — | 25 Sep 2026 | Arena |
Other gpt-oss versions
- gpt-oss-safeguard-120b29 Oct 2025
- gpt-oss-safeguard-20b29 Oct 2025
- gpt-oss-120b5 Aug 2025