Artificial Analysis profiles open-weights gpt-oss-20b
Artificial Analysis has published a benchmark page for gpt-oss-20b, a 21B-parameter Apache 2.0 reasoning model with 3.6B active parameters, so builders can compare a small open-weights model on its intelligence, speed, and price.
Benchmarking firm Artificial Analysis has published a profile of gpt-oss-20b. It compares two variants of the model's release on intelligence, performance, and price.
According to the page, gpt-oss-20b is a reasoning model with 21 billion total parameters, of which 3.6 billion are active per token during inference. It takes text in and produces text out. Its context window is 131k tokens, and its knowledge cutoff is May 31, 2024. The weights are on Hugging Face under an Apache 2.0 license.
On latency, Artificial Analysis says the low-reasoning-effort variant, gpt-oss-20b (Low), has the shortest time to first answer token of the two, at 2.23 seconds.
The firm groups open-weights models by size when it compares them. At 21B parameters, gpt-oss-20b is in the "Small" class (4B–40B parameters), so it is ranked against other open-weights models of that size rather than against larger frontier systems. Because it is a reasoning model, it is compared with both reasoning and non-reasoning models.
The page lists the evaluations behind its capability indexes. They include AA-Omniscience v1.1, AA-Briefcase v1.1, HLE-Diamond, AutomationBench-AA, GDP.pdf, Terminal-Bench 4.0, FrontierCode-AA 2.0 and CritPt v2 (Hard Subset). The material available does not include the model's actual index scores, cost per task, or pricing.