This model is 13 months old. A newer version of it has since shipped: GPT-6 Astra (2026-09-03). The details below are still accurate for this model — it just is not what we would recommend today.
gpt-oss-20b
OpenAI🇺🇸 United States
Smaller open-weights sibling of gpt-oss-120b: 21B-param MoE (3.6B active), Apache 2.0, runs in 16GB of memory for on-device and edge use.
Context window131K tokens
Input / 1M tokens$0.03
Output / 1M tokens$0.13
Version History
gpt-oss-20b-launchmajor
gpt-oss-20b launches. Smaller open-weights sibling of gpt-oss-120b: 21B-param MoE (3.6B active), Apache 2.0, runs in 16GB of memory for on-device and edge use.
Benchmark Scores
Full leaderboard →89.3%
AIME 2025
68.8%
GPQA
86.7%
Hallucination Rate
77.7%
LiveCodeBench
85.3%
MMLU
74.8%
MMLU-Pro
192.0 tokens_per_sec
Speed (tok/s)