Wait Which Model?
← Back to directory
OpenAI

gpt-oss-20b

Released Aug 5, 2025 · knowledge cutoff 2024-06

Status
Superseded
Location
United States
Modality
Text
Context window
131K
Max output
131K
Speed (high effort)
151 tok/s · 0.88s to first answer token
Price ($/MTok in / out)
$0.03 / $0.13
Cost per task
—
Open weights
Yes
Benchmarks5 of 8 reported

MMLU-Pro

73.6%

GPQA Diamond

71.5%

SWE-bench Verified

60.7%

AIME (latest)

91.7%

Humanity's Last Exam

10.9%

Terminal-Bench 2.1

—

LMArena Elo

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Runs in about 16GB of memory, putting it within reach of a single consumer GPU or a high-end laptop
  • Competition math performance keeps pace with its 120B sibling despite the much smaller footprint
  • Same effort-string reasoning control and exposed chain-of-thought as the 120B model

Weaknesses

  • Hallucinates even more than the 120B on people and fact lookups, the cost of a smaller knowledge store
  • Also requires the Harmony response format — a generic chat template silently breaks tool calls
  • Agentic terminal/tool-use tasks lag well behind its math and coding strength — a narrow specialist rather than an all-rounder

Released alongside gpt-oss-120b as the lower-latency, edge-deployable sibling. GPQA Diamond 71.5%, HLE 10.9% and AIME 2025 91.7% are no-tools scores at high reasoning effort per OpenAI's model card (arXiv:2508.10925); with tools, GPQA rises to 74.2%, HLE to 17.3%, AIME to 98.7% (edging past the 120B model). MMLU-Pro 73.6 is from the official Hugging Face model card. Terminal-Bench reported is 2.0 (roughly 3%, sources conflict between 3.1% and 3.4%), not the tracked 2.1 metric, so left null here. Pricing ($0.03/$0.13 per MTok) is a representative third-party hosted rate (OpenRouter) — OpenAI's own API docs list the model with no attached price, and rates vary across hosts (e.g. AWS Bedrock charges $0.07/$0.30).

For developers

API model strings

  • openaigpt-oss-20b
  • aws-bedrockopenai.gpt-oss-20b-1:0

Licence

Apache License 2.0 · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor

News