OpenAIproprietary
Released: 2026-06-01OpenAI: gpt-oss-20b
gpt-oss-20b is an open-weight 21B parameter model released by OpenAI under the Apache 2.0 license. It uses a Mixture-of-Experts (MoE) architecture with 3.6B active parameters per forward pass, optimized for...
Intelligence Index
77.0/ 100
Calibrated multi-domain compositeThroughput Speed
130tok/sec
Output streaming throughputTime To First Token
280ms
Initial chunk server latencyLiveBench Contam-Free
67.8%
September 2026 suite scoreLiveBench Multi-Domain Evaluation Breakdown
Monthly refreshed contamination-free test setsLogical Reasoning70.8%
Coding & Repo Repair69.3%
Mathematics (AIME)68.5%
Data Analysis & Tables65.5%
Language & Comprehension67.0%
Instruction Following67.8%
Sourced Benchmark Evaluations (61)
Verified performance across authoritative benchmarks with provenance tracking.| Benchmark | Status | Score | Trust | Date | Source Type | Provenance |
|---|---|---|---|---|---|---|
| ARC-AGI-3 | Active | 48.6% | 93 | 2026-07-06 | independent | Source ↗ |
| ARC-AGI-2 | Nearing Saturation | 58.3% | 75 | 2026-06-29 | independent | Source ↗ |
| FrontierMath | Active | 20% | 98 | 2026-06-26 | independent | Source ↗ |
| PutnamBench | Active | 37.2% | 90 | 2026-06-25 | independent | Source ↗ |
| SWE-Lancer | Deprecated | 49.7% | 20 | 2026-06-23 | independent | Source ↗ |
| Omni-MATH | Deprecated | 53% | 10 | 2026-06-23 | independent | Source ↗ |
| Soohak | Active | 54.4% | 98 | 2026-06-21 | independent | Source ↗ |
| FrontierCode | Active | 42.4% | 98 | 2026-06-21 | independent | Source ↗ |
| Humanity's Last Exam | Active | 32.9% | 93 | 2026-06-19 | independent | Source ↗ |
| StructFlowBench | Nearing Saturation | 57.6% | 75 | 2026-06-19 | independent | Source ↗ |
| AIME | Nearing Saturation | 68% | 75 | 2026-06-19 | independent | Source ↗ |
| OlympiadBench | Active | 58.7% | 80 | 2026-06-18 | independent | Source ↗ |
| MathArena | Active | 57.3% | 99 | 2026-06-17 | independent | Source ↗ |
| FireBench | Active | 57% | 95 | 2026-06-17 | independent | Source ↗ |
| Terminal-Bench | Active | 51.2% | 80 | 2026-06-17 | independent | Source ↗ |
| OpenAI-MRCR | Active | 64.7% | 80 | 2026-06-16 | independent | Source ↗ |
| SimpleQA | Active | 40% | 90 | 2026-06-16 | independent | Source ↗ |
| CodeContests | Deprecated | 47.5% | 10 | 2026-06-16 | independent | Source ↗ |
| ComplexBench | Nearing Saturation | 55.6% | 80 | 2026-06-15 | independent | Source ↗ |
| SWE-bench Pro | Active | 45.3% | 90 | 2026-06-15 | independent | Source ↗ |
| LongBench-Pro | Active | 63.1% | 80 | 2026-06-15 | independent | Source ↗ |
| ZeroSCROLLS | Nearing Saturation | 61.6% | 78 | 2026-06-14 | independent | Source ↗ |
| MultiPL-E | Deprecated | 67.3% | 10 | 2026-06-14 | independent | Source ↗ |
| LiveCodeBench | Saturated | 59.9% | 39 | 2026-06-13 | independent | Source ↗ |
| AgentIF | Nearing Saturation | 46.1% | 80 | 2026-06-13 | independent | Source ↗ |
| MATH | Saturated | 70.2% | 25 | 2026-06-13 | independent | Source ↗ |
| InfiniteBench | Nearing Saturation | 65.5% | 65 | 2026-06-13 | independent | Source ↗ |
| Mercury | Deprecated | 64.3% | 10 | 2026-06-12 | independent | Source ↗ |
| SCROLLS | Nearing Saturation | 61.6% | 65 | 2026-06-12 | independent | Source ↗ |
| MMLU-Pro | Nearing Saturation | 70.8% | 75 | 2026-06-12 | independent | Source ↗ |
| RULER | Nearing Saturation | 70.8% | 84 | 2026-06-11 | independent | Source ↗ |
| TabMWP | Saturated | 65.9% | 25 | 2026-06-11 | independent | Source ↗ |
| CyberSecEval | Deprecated | 65.8% | 10 | 2026-06-11 | independent | Source ↗ |
| SWE-bench Verified | Saturated | 57% | 25 | 2026-06-10 | independent | Source ↗ |
| MMLU-Redux | Saturated | 75.5% | 25 | 2026-06-10 | independent | Source ↗ |
| QuALITY | Nearing Saturation | 67.8% | 78 | 2026-06-10 | independent | Source ↗ |
| LongBench | Nearing Saturation | 66.2% | 65 | 2026-06-09 | independent | Source ↗ |
| SWE-bench | Deprecated | 49.7% | 10 | 2026-06-09 | independent | Source ↗ |
| MGSM | Nearing Saturation | 68% | 65 | 2026-06-09 | independent | Source ↗ |
| AGIEval | Saturated | 64.2% | 25 | 2026-06-08 | independent | Source ↗ |
| QASPER | Nearing Saturation | 63.9% | 65 | 2026-06-08 | independent | Source ↗ |
| TruthfulQA | Saturated | 67.8% | 25 | 2026-06-08 | independent | Source ↗ |
| GSM-Hard | Deprecated | 68.7% | 10 | 2026-06-07 | independent | Source ↗ |
| SuperGLUE | Saturated | 76.2% | 25 | 2026-06-07 | independent | Source ↗ |
| CodeXGLUE | Deprecated | 68.7% | 23 | 2026-06-06 | independent | Source ↗ |
| MT-Bench | Nearing Saturation | 7% | 75 | 2026-06-06 | independent | Source ↗ |
| BIG-Bench Hard | Saturated | 73.2% | 25 | 2026-06-06 | independent | Source ↗ |
| WinoGrande | Saturated | 77.7% | 28 | 2026-06-05 | independent | Source ↗ |
| MBPP+ | Nearing Saturation | 74.6% | 65 | 2026-06-05 | independent | Source ↗ |
| IFEval | Deprecated | 66.4% | 25 | 2026-06-05 | independent | Source ↗ |
| GPQA | Nearing Saturation | 67.8% | 78 | 2026-06-05 | independent | Source ↗ |
| Needle-in-a-Haystack | Saturated | 95.4% | 44 | 2026-06-04 | independent | Source ↗ |
| HumanEval+ | Saturated | 78.9% | 25 | 2026-06-04 | independent | Source ↗ |
| CommonsenseQA | Saturated | 78.4% | 28 | 2026-06-04 | independent | Source ↗ |
| MBPP | Saturated | 77.5% | 25 | 2026-06-03 | independent | Source ↗ |
| GSM8K | Saturated | 77.3% | 25 | 2026-06-03 | independent | Source ↗ |
| GLUE | Saturated | 80.7% | 28 | 2026-06-03 | independent | Source ↗ |
| MMLU | Saturated | 78.5% | 25 | 2026-06-02 | independent | Source ↗ |
| HumanEval | Saturated | 83.3% | 25 | 2026-06-02 | independent | Source ↗ |
| ARC-AGI | Saturated | 70.2% | 25 | 2026-06-02 | independent | Source ↗ |
| LiveBench (Continuous) | Active | 67.8% | 94 | 2025-02-01 | independent | Source ↗ |
OpenAI Compatible API ExecutionAPI Ready
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="openai-gpt-oss-20b",
messages=[{"role": "user", "content": "Evaluate multi-step logic problem"}],
temperature=0.2
)
print(response.choices[0].message.content)Evaluation & Sourcing Notes
Scores listed for OpenAI: gpt-oss-20b represent verified evaluations extracted from official research papers, independent evaluation suites (HELM, LMSYS, OpenCompass, LiveBench), and verified audit reports.
All benchmarks marked as saturated or deprecated reflect historical performance where the benchmark no longer provides active discriminative power.