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.
Added Aug 5, 2025
Model weightsContext Window
128.0K
Max Output
16.4K
Input Price (Auto)
$0.030/1M
Output Price (Auto)
$0.13/1M
Cache Read (Auto)
$0.030/1M
Capabilities
Benchmarks
Benchmarks
Performance metrics and benchmarks
Sourced from Artificial Analysis.
Intelligence Index
15.2
Coding Index
20.7
Agentic Index
3.1
Reasoning
GPQA Diamond
Graduate-level scientific reasoning
68.8%
Better than 49% of models compared
HLE
Humanity's Last Exam
11.0%
Better than 60% of models compared
IFBench
Instruction-following benchmark
65.1%
Better than 77% of models compared
T²-Bench Telecom
Conversational AI agents in dual-control scenarios
60.2%
Better than 57% of models compared
AA-LCR
Long context reasoning evaluation
33.3%
Better than 40% of models compared
GDPval-AA
Economically valuable tasks
3.2%
CritPt
Research-level physics reasoning
1.4%
Coding
SciCode
Python programming for scientific computing
34.4%
Better than 51% of models compared
Terminal-Bench Hard
Agentic coding and terminal use
10.6%
Better than 45% 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%
Last updated Aug 16, 2026
Artificial AnalysisProviders
Choose explicit providers for this model. Auto routing remains available as the default option.
Loading provider options…
Related text models
Compare GPT OSS 20B with similar models from the same provider or model family.
GPT OSS Safeguard 20B
openai/gpt-oss-safeguard-20bgpt-oss-safeguard is a first open weight reasoning model specifically trained for safety classification tasks to help classify text content based on customizable policies. As a fine-tuned version of gpt-oss, gpt-oss-safeguard is designed to follow explicit written policies that you provide. This enables bring-your-own-policy Trust & Safety AI, where your own taxonomy, definitions, and thresholds guide classification decisions. Well crafted policies unlock gpt-oss-safeguard's reasoning capabilities, enabling it to handle nuanced content, explain borderline decisions, and adapt to contextual factors.
GPT OSS 120B
openai/gpt-oss-120bAn open-weight, 117B-parameter Mixture-of-Experts (MoE) language model designed for high-reasoning, agentic, and general-purpose production use cases. It activates 5.1B parameters per forward pass and is optimized to run on a single H100 GPU with native MXFP4 quantization. The model supports configurable reasoning depth, full chain-of-thought access, and native tool use, including function calling, browsing, and structured output generation.
GPT 5.6 Luna
openai/gpt-5.6-lunaGPT-5.6 Luna is the fast, cost-efficient model in OpenAI's GPT-5.6 series. It is suited for high-volume chat, classification, lightweight agentic workflows, and latency-sensitive reasoning tasks.
GPT 5.6 Luna Pro
openai/gpt-5.6-luna-proGPT-5.6 Luna with Pro reasoning mode. Pro mode performs more model work for difficult tasks; reasoning effort remains independently configurable and defaults to medium.
GPT 5.6 Sol
openai/gpt-5.6-solGPT-5.6 Sol is the flagship model in OpenAI's GPT-5.6 series. It is suited for complex reasoning, coding, agentic workflows, command-line work, and multi-step professional tasks.
GPT 5.6 Sol Pro
openai/gpt-5.6-sol-proGPT-5.6 Sol with Pro reasoning mode. Pro mode performs more model work for difficult tasks; reasoning effort remains independently configurable and defaults to medium.