LLMOpen weights

gpt-oss-20b

OpenAI

Updated May 21, 2026 · Benchmarks via Artificial Analysis, specs & pricing via OpenRouter · Methodology · Report an error

Specifications

Type: LLM
Access: Open weights
Released: August 5, 2025
License: Apache 2.0
Parameters: 21B (3.6B active)
Context window: 131,072 tokens
Knowledge cutoff: June 30, 2024
Output speed: 1000 tok/s
Latency (TTFT): 0.38s
Input: Text
Output: Text
API pricing: $0.03 in · $0.14 out / 1M tokens

Capabilities

Function callingStructured outputCode execution

Benchmarks

Reasoning

71.5

Coding

77.7

Math

89.3

Agents

TAU-bench Retail

54.8

General

Humanity’s Last Exam	17.3
MMLU	85.3
MMLU-Pro	74.8

Our take

Edge-deployable sibling of gpt-oss-120b (~16GB RAM), roughly o3-mini-level on common benchmarks.

Related research

gpt-oss-120b & gpt-oss-20b Model CardModel paper

Links

Model card Announcement

Compare gpt-oss-20b with

vs Sonar Reasoning Pro vs R1 1776 vs Qwen3.7 Max vs Gemini 3.5 Flash

See all gpt-oss-20b alternatives →