gpt-oss-20b: specs and benchmarks

Key specs

Provider: OpenAI
Family: gpt-oss
API identifier: gpt-oss-20b
Availability: Available
Release date: 2025-08-05
License: Open source (Apache 2.0)
Context window: 131,072 tokens
Max output: 131,072 tokens
Knowledge cutoff: June 2024
API access: Unknown
Consumer app: Unknown

Capabilities

Inputs

Text input Yes
Image input No Model page marks image as not supported; text is the only modality
Audio input No Model page marks audio as not supported
Video input No Model page marks video as not supported
File input Unknown

Outputs

Text output Yes
Image output No Model page marks image as not supported; text is the only modality
Audio output No Model page marks audio as not supported

Features

Reasoning Native Configurable reasoning effort (low, medium, high) with full chain-of-thought access; no option to disable reasoning is documented
Tool calling Yes Function calling listed as a supported feature and Responses API tool; key features name native function calling, web browsing, Python code execution and structured outputs
Web search Yes Model page lists web_search among the tools supported with the Responses API (with function_calling, code_interpreter and mcp); key features also name native web browsing

"Unknown" means we have not verified this capability yet; it does not mean the model lacks it.

Pricing

No current prices listed.

Benchmarks

Benchmark	Setting	Score	Reported by	Measured	Verified	Source
GDPval-AA v2.1 Elo, higher is better	high effort	339 Elo	Third party	2026-10-25	2026-10-25	GDPval-AA v2.1 Leaderboard (Artificial Analysis)
Humanity's Last Exam accuracy %, higher is better	no tools; high effort	10.9 %	Self-reported	—	2026-10-25	gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, arXiv 2508.10925)
	no tools; low effort	4.2 %	Self-reported	—	2026-10-25	gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, arXiv 2508.10925)
	no tools; medium effort	7 %	Self-reported	—	2026-10-25	gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, arXiv 2508.10925)
	text-only subset; Scale Labs harness, o3-mini judge	9.73 %	Third party	—	2026-10-25	Humanity's Last Exam (Text Only) leaderboard (Scale Labs)
	with tools; high effort	17.3 %	Self-reported	—	2026-10-25	gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, arXiv 2508.10925)
	with tools; low effort	6.3 %	Self-reported	—	2026-10-25	gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, arXiv 2508.10925)
	with tools; medium effort	8.8 %	Self-reported	—	2026-10-25	gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, arXiv 2508.10925)
Text Arena Arena score, higher is better	Overall, style control on	1,317	Third party	2026-10-23	2026-10-25	Text Arena leaderboard (Arena)

GDPval-AA v2.1 caveats:

Elo is a relative rating: values change as models are added and the scale is re-anchored, so always store the date a rating was read. For example, Anthropic's launch page gives 1846 for Claude Opus 5.5, while the leaderboard on 2026-10-12 shows 1866 (Max, Default Fallback). Earlier versions (Google cites "GDPVal-AA v2") are not comparable. Judging is done by an LLM, not by human experts. Anthropic models are listed with "Default Fallback", meaning refused steps are routed to a fallback model. AA-Briefcase and GDP.pdf are separate evaluations.

Humanity's Last Exam caveats:

Results differ a lot by setting: with tools versus without tools (Anthropic reports 45.9% without tools and 57.4% with tools for Claude Haiku 5.5), and the full set versus the text-only subset. Providers use different judge models and harnesses (Scale uses o3-mini; Alibaba states GPT-4o). Derived sets such as HLE-Verified (a third-party revision with 1,811 scored items), HLE-Rolling and HLE-Diamond are different item sets and are not comparable. The questions are public, so contamination is possible, and community reviews have reported flawed items.

Text Arena caveats:

Scores measure voter preference, which can be influenced by answer style and length. The leaderboard offers a style-control adjustment, so record which view was used. Scores shift with every update, so a value is only meaningful together with its leaderboard date. Entries marked "Preliminary" or with few votes have wide confidence intervals. The same model can appear several times with different reasoning-effort suffixes (for example -high or -max). Category views (Coding, Math, Creative Writing and others) are different rankings from Overall.

One row per result, without rank numbers: each result was measured under its own setting (effort level, harness, tools, test version), as stated by the source, and scores compare only within the same setting. Self-reported results come from the provider and have not been reproduced independently.

Rankings

#14 of 14 in Best AI for research as of 2026-10-26
#42 of 42 in Best AI API as of 2026-10-26

Derived from our published ranking snapshots; each ranking explains its criteria and weights.

Compare gpt-oss-20b

Compare with

Guides

About gpt-oss-20b

gpt-oss-20b is OpenAI's medium-sized open-weight model for low-latency, local or specialized use cases, with 21B total parameters and 3.6B active parameters. It is released under the Apache 2.0 license with weights available on Hugging Face, and OpenAI's announcement states it can run on edge devices with 16 GB of memory. It supports configurable reasoning effort (low, medium, high), full chain-of-thought access, function calling and structured outputs, and OpenAI describes it as fine-tunable.

Strengths and weaknesses

Strengths

Open weights under the Apache 2.0 license; can be self-hosted and fine-tuned
Can run on edge devices with 16 GB of memory, according to OpenAI's announcement
Configurable reasoning effort (low, medium, high) with full chain-of-thought access
Native function calling, web browsing, Python code execution and structured outputs

Weaknesses

Text-only input and output (no image input)
131,072-token context window and June 2024 knowledge cutoff, smaller and earlier than the GPT-6 models
No OpenAI API price is listed on the pricing page, and the model page shows rate limits of 0 for all API usage tiers