ModelBeat
Models/gpt-oss-20b
All models
O

OpenAI: gpt-oss-20b

Watch this modelCompareAnnouncement
openai/gpt-oss-20b

gpt-oss-20b is OpenAI’s open-weights AI model, released Aug 5, 2025. It costs $0.02/M input and $0.09/M output tokens, and handles a 131K-token context window.

gpt-oss-20b is an open-weight 21B parameter model released by OpenAI under the Apache 2.0 license.

Context
131K
Input
$0.02 / 1M
Output
$0.09 / 1M
Released
Aug 5, 2025
Modalities
text→text

Benchmarks

Scores on standardized evaluations. Higher is better — and rank shows where gpt-oss-20b lands among all models tracked on Model Beat.

7.0
Intelligence Index
Epoch AI
7th percentile of tracked models
10.0
Coding Index
Epoch AI
10th percentile of tracked models
6.0
Agentic Index
Epoch AI
6th percentile of tracked models

Reasoning

4 evals
GPQA Diamond60.8%

Graduate-level scientific reasoning

Humanity's Last Exam5.3%

Frontier of human expert knowledge

MMLU-Pro71.8%

Broad expert knowledge

WeirdML40.9%

Novel ML problem-solving

Coding

2 evals
LiveCodeBench65.2%

Contamination-free coding

SciCode38.9%

Scientific research coding

Math

1 evals
AIME 2024/202565.3%

Olympiad-qualifier math

Agentic & Tools

2 evals
Terminal-Bench3.4%

Command-line agentic tasks

τ²-bench50.3%

Tool-agent-user reliability

In the news

1

Changelog

11
  • gpt-oss-20b: cheapest credible provider via Darkbloom now $0.09 per 1M output tokens (was $0.1, -10%)cheapest provider · Sep 22, 2026
  • gpt-oss-20b: cheapest credible provider via Darkbloom now $0.018 per 1M input tokens (was $0.02, -10%)cheapest provider · Sep 22, 2026
  • gpt-oss-20b: reported LiveCodeBench score fell from 77.7% to 65.2% (-16%)benchmark · Sep 8, 2026
  • gpt-oss-20b: reported τ²-bench score fell from 60.2% to 50.3% (-17%)benchmark · Sep 8, 2026
  • gpt-oss-20b: reported Humanity's Last Exam score fell from 11.0% to 5.3% (-52%)benchmark · Sep 8, 2026
  • gpt-oss-20b: reported SciCode score rose from 34.4% to 38.9% (+13%)benchmark · Sep 7, 2026
  • gpt-oss-20b: cheapest credible provider via Darkbloom now $0.1 per 1M output tokens (was $0.14, -29%)cheapest provider · Aug 28, 2026
  • gpt-oss-20b: cheapest credible provider via Darkbloom now $0.02 per 1M input tokens (was $0.03, -33%)cheapest provider · Aug 28, 2026
  • gpt-oss-20b: reported AIME 2024/2025 score rose from 53.9% to 65.3% (+21%)benchmark · Aug 27, 2026
  • gpt-oss-20b: reported GPQA Diamond score rose from 46.0% to 60.8% (+32%)benchmark · Aug 27, 2026
  • gpt-oss-20b: Humanity's Last Exam improved from 9.8% to 11.0% (+12%)benchmark · Aug 6, 2026

Compare gpt-oss-20b with

gpt-oss-20b vs GPT-6 Astragpt-oss-20b vs Claude Fable 5gpt-oss-20b vs Kimi K3gpt-oss-20b vs Gemini 3.7 Flashgpt-oss-20b vs Grok 4.6

Frequently asked questions

What is gpt-oss-20b?

gpt-oss-20b is an open-weight 21B parameter model released by OpenAI under the Apache 2.0 license.

Who created gpt-oss-20b?

gpt-oss-20b is an AI model developed by OpenAI, part of the GPT family, released Aug 5, 2025.

How much does gpt-oss-20b cost?

gpt-oss-20b is served from $0.02 per million input tokens and $0.09 per million output tokens (cheapest credible provider on OpenRouter: Darkbloom).

What is gpt-oss-20b's context window?

gpt-oss-20b supports a context window of up to 131K tokens.

How does gpt-oss-20b perform on benchmarks?

On standardized evaluations tracked by Epoch AI, gpt-oss-20b scores 38.9% on SciCode, 40.9% on WeirdML, 65.2% on LiveCodeBench.

Is gpt-oss-20b open source?

gpt-oss-20b is released with open weights (Open weights (unrestricted)), so the model can be downloaded and self-hosted.

Benchmark data from Epoch AI (CC BY) and Artificial Analysis; pricing, specs & descriptions from OpenRouter. Data updated Oct 6, 2026.