gpt-oss-120b
Released Aug 5, 2025 · knowledge cutoff 2024-06
- Status
- Superseded
- Location
- United States
- Modality
- Text
- Context window
- 131K
- Max output
- 131K
- Speed (high effort)
- 170 tok/s · 0.88s to first answer token
- Price ($/MTok in / out)
- $0.03 / $0.17
- Cost per task
- —
- Open weights
- Yes
Benchmarks5 of 8 reported
MMLU-Pro
GPQA Diamond
SWE-bench Verified
AIME (latest)
Humanity's Last Exam
Terminal-Bench 2.1
LMArena Elo
ARC-AGI-2
Strengths & weaknesses
Strengths
- Fits on a single 80GB datacenter GPU at native precision, so it's a genuinely self-hostable release rather than a paper-only open weights drop
- Reasoning effort is a plain system-prompt string (low/medium/high), trading latency for depth without swapping checkpoints
- Exposes its full chain-of-thought in the output rather than summarizing or hiding it, unlike OpenAI's closed reasoning models
Weaknesses
- Hallucinates markedly more than OpenAI's closed models on people and fact lookups — the expected trade-off of a smaller open weight store of world knowledge
- Locked to OpenAI's bespoke Harmony response format; feeding it a generic chat template silently breaks tool calls and reasoning
- Inherited safety training leans toward refusal, and self-hosters report it overcorrecting on benign prompts
OpenAI's first open-weight release since GPT-2. GPQA Diamond 80.1%, HLE 14.9% and AIME 2025 92.5% are no-tools scores at high reasoning effort per OpenAI's model card (arXiv:2508.10925); with tools, GPQA rises to 80.9%, HLE to 19.0%, AIME to 97.9%. MMLU-Pro 80.8 is from the official Hugging Face model card. Terminal-Bench reported is 2.0 (18.7%), not the tracked 2.1 metric, so left null here. Pricing ($0.03/$0.17 per MTok) is a representative third-party hosted rate (OpenRouter/CoreWeave) — OpenAI's own API docs list the model with no attached price, and rates vary roughly 9x across hosts (e.g. AWS Bedrock charges $0.15/$0.60).
For developers
API model strings
- openai
gpt-oss-120b - aws-bedrock
openai.gpt-oss-120b-1:0
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor