What it is
gpt-oss-120b is the larger of OpenAI's two open-weight models, released in August 2025. It has 117 billion parameters but uses a mixture-of-experts design, so only 5.1 billion are active for each token. That, plus 4-bit (MXFP4) weights, lets it run on one 80GB GPU such as an NVIDIA H100 or AMD MI300X. It has a 128K-token context.
It is a text-only reasoning model with low, medium and high effort settings, set in the system prompt, and built-in support for function calling, web browsing, Python execution and structured outputs. It must be prompted in OpenAI's harmony format, and its full chain of thought is exposed for debugging but not meant for end users.
The weights are free under Apache 2.0. You run it yourself or through hosts such as Amazon Bedrock, OpenRouter and Ollama's cloud. As of October 2026 OpenAI has not released a successor; its gpt-oss-safeguard policy models are a separate release.
Our review
gpt-oss-120b has aged into a utility: not the open model with the best scores, but the one with the fewest strings and the most tooling. Artificial Analysis gives its high-effort setting 12 on its Intelligence Index, 11th of 67 open-weight models between 40B and 150B parameters, where the median is 8. IBM's dense Granite 4.2 30B scores 13 on the same index with about a quarter of the parameters.
OpenAI's own numbers were strong at launch: 62.4% on SWE-bench Verified at high effort and a GPQA Diamond score of about 80, all run on the MXFP4 weights. Those are self-reported. The model card's harder figure is 16.2 on SWE-Bench Pro from Scale's leaderboard, which is closer to how it handles real repository work today.
The economics are what keep it in use. With 5.1B parameters active per token it is fast: Artificial Analysis records a median of 225.5 tokens a second across providers, fourth of 27 in its class, at $0.15 per million input tokens and $0.60 per million output. Self-hosted, the Ollama build is a 65GB download that fits one H100 with room for context.
The friction is in the format. The card says the model "will not work correctly" outside OpenAI's harmony format; the Transformers chat template applies it, but code that calls generate directly or a serving stack with an old template gets degraded output. Context stops at 128K, there is no image input, and fine-tuning the 120b needs a full H100 node where the 20b fits consumer hardware.
What still sets it apart from the ten models above it in its class is the ecosystem. Hugging Face counted almost 3 million downloads in the last month, 124 fine-tunes and 135 quantisations. A compliance team can approve Apache 2.0 from an American lab quickly, and whichever quantisation an engineer needs very likely exists already.
Where it earns its place
A platform engineer with one H100 in a private cloud
Serve an internal assistant that never sends data to a third party
vllm serve or Ollama runs the MXFP4 weights on the single 80GB card; reasoning effort is set per request in the system prompt. The cost is the GPU itself, with no per-token bill and nothing leaving the network.
A support-operations lead at a fintech handling 300,000 tickets a month
Draft replies and tag tickets cheaply through a hosted endpoint
At 2,000 input and 500 output tokens a ticket and the $0.15/$0.60 per million rates Artificial Analysis lists, the month costs about $180. Batch and Flex tiers on Bedrock halve that for overnight work.
An ML team at a legal-tech company
Fine-tune a reasoning model on its own contracts and redistribute it to clients
Apache 2.0 permits both, and fine-tuning the 120b needs one H100 node. The tuned model keeps the harmony format and 128K context, so long contract bundles still need chunking.
Before you commit
- Setup
- ollama pull gpt-oss:120b (a 65GB download), lms get in LM Studio, or vllm serve; set Reasoning: low, medium or high in the system prompt.
- Integrations
- Transformers, vLLM, SGLang, Ollama, LM Studio and Docker Model Runner locally; Amazon Bedrock, OpenRouter and other hosts through OpenAI-compatible APIs.
- Data and privacy
- Apache 2.0 with no usage restrictions in the licence. Self-hosted, prompts stay on your hardware; hosted, retention follows each provider's terms, and OpenRouter's EU-only routing needs a Business or Enterprise plan.
- Cost at scale
- 1B input and 250M output tokens a month cost about $300 at $0.15/$0.60 per million; Bedrock batch halves its rate. Self-hosting needs one 80GB GPU for inference and an H100 node to fine-tune.
- Limits
- 128K context, text only; harmony format required; raw chain of thought should be hidden from users; Artificial Analysis ranks it 11th of 67 in its class.
What's good
- Apache 2.0 with no usage restrictions in the licence
- Fits one 80GB H100 or MI300X in MXFP4
- About 225 tokens a second median across hosts
- 124 fine-tunes and 135 quantisations on Hugging Face
What's not
- Ten open models in its size class score higher on Artificial Analysis
- 128K context and text only
- Breaks quietly outside the harmony prompt format
- 16.2 on SWE-Bench Pro
Best for
Teams that need a permissively licensed reasoning model they can run privately on one GPU, with mature tooling.
Alternatives
gpt-oss-120b: common questions
- Is gpt-oss-120b worth it?
- Fourteen months on, gpt-oss-120b is no longer the strongest open model, but it is still the plainest licence and the widest support for reasoning on one 80GB GPU. We rate it 4.0 out of 5.
- How much does gpt-oss-120b cost?
- Free weights under Apache 2.0; hosted prices vary by provider.
- What are the downsides of gpt-oss-120b?
- Ten open models in its size class score higher on Artificial Analysis; 128K context and text only; breaks quietly outside the harmony prompt format; 16.2 on SWE-Bench Pro.
- Who is gpt-oss-120b best for?
- Teams that need a permissively licensed reasoning model they can run privately on one GPU, with mature tooling.
- How much does gpt-oss-120b cost for a team or in production?
- 1B input and 250M output tokens a month cost about $300 at $0.15/$0.60 per million; Bedrock batch halves its rate. Self-hosting needs one 80GB GPU for inference and an H100 node to fine-tune.
- Is gpt-oss-120b safe for company data?
- Apache 2.0 with no usage restrictions in the licence. Self-hosted, prompts stay on your hardware; hosted, retention follows each provider's terms, and OpenRouter's EU-only routing needs a Business or Enterprise plan.
- What are the best gpt-oss-120b alternatives?
- The closest alternatives we have reviewed are gpt-oss-20b, DeepSeek-V4-Pro, Qwen3.8-Max and Kimi K3.
Answers checked Oct 22, 2026.
Make gpt-oss-120b? Show this rating on your site
