GPT OSS 20B

An open-weight 21B parameter model released under the Apache 2.0 license. It uses a Mixture-of-Experts (MoE) architecture with 3.6B active parameters per forward pass, optimized for lower-latency inference and deployability on consumer or single-GPU hardware. The model is trained in OpenAI's Harmony response format and supports reasoning level configuration, fine-tuning, and agentic capabilities including function calling, tool use, and structured outputs.

Reasoning

Added Aug 5, 2025

Model weights

Pricing

Auto routing · per 1M tokens

Input: $0.020
Output: $0.100

Compare provider prices

Specifications

Context window: 128K
Max output: 16.4K
Parameters: 20B / 3.6B; Total / active
Avg output (7d): 3K tokens; Longer than 97% of models

Benchmarks

Artificial Analysis

LMArena

Design Arena

Sourced from Artificial Analysis.

Intelligence Index

9.0

Better than 35% of models compared

Coding Index

20.7

Better than 23% of models compared

Agentic Index

1.2

Better than 22% of models compared

Agentic work

AutomationBench-AA

Workflow automation with guardrail penalties

0.2%

Better than 1% of models compared

AA-Briefcase

Agentic knowledge work (Elo)

0 Elo

Better than 3% of models compared

GDPval-AA v2

Economically valuable tasks (Elo)

343 Elo

Better than 20% of models compared

Document reasoning

GDP.pdf

Professional PDF reasoning: all-pass rate

2.0%

Better than 13% of models compared

AA-LCR v1.1

Long context reasoning with updated grading

34.7%

Better than 33% of models compared

MLCR-AA

Medical long-context reasoning

0.6%

Better than 4% of models compared

Reasoning

HLE

Humanity's Last Exam

11.0%

Better than 53% of models compared

IFBench

Instruction-following benchmark

65.1%

Better than 77% of models compared

CritPt

Research-level physics reasoning

1.4%

Coding

Terminal-Bench v4.0

Practical coding and terminal tasks

0.0%

Better than 16% of models compared

SciCode

Python programming for scientific computing

38.9%

Better than 22% of models compared

LiveCodeBench

Contamination-free coding benchmark

77.7%

Better than 89% of models compared

Math

AIME 2025

American Invitational Mathematics Examination 2025

89.3%

Better than 88% of models compared

Knowledge

MMLU-Pro

Professional and academic subject knowledge

74.8%

Better than 47% of models compared

AA-Omniscience Accuracy

Proportion of correctly answered questions

16.0%

AA-Omniscience Hallucination Rate

Rate of incorrect answers among non-correct responses

94.1%

Legacy benchmarks

GPQA Diamond (legacy)

Graduate-level scientific reasoning

68.8%

Better than 46% of models compared

Terminal-Bench Hard (legacy)

Agentic coding and terminal use

10.6%

Better than 45% of models compared

T²-Bench Telecom (legacy)

Conversational AI agents in dual-control scenarios

60.2%

Better than 57% of models compared

AA-LCR (unversioned / legacy)

Long context reasoning evaluation

34.7%

Better than 33% of models compared

GDPval-AA (unversioned / legacy)

Economically valuable tasks