Chat with GPT-OSS 120B.

OpenAI’s open-weight model. Only 5B of its 117B works on each word, so it answers fast, and it is built to reason step by step.

Start chatting with GPT-OSS 120B now — free, no account

At a glance

  • Maker: OpenAI
  • Released: 2025
  • Parameters: 117B
  • Active per token: 5B
  • Context window: 128k
  • Size: Large
  • Memory: Handles long documents
  • Good at: Reasoning and agent work
  • Supports: tools, zero data retention
  • License: Apache 2.0
  • Upstream API price: $0.60 per million tokens

Announced in Introducing gpt-oss.

What is GPT-OSS 120B?

GPT-OSS 120B is OpenAI's open-weight model — the company's first genuinely open release since the early days, and proof that the frontier labs can play the open game when they choose to. It is a reasoning model, built to work problems through step by step rather than answer off the top of its head, and its Apache-2.0 licence is about as permissive as licences get: run it, fine-tune it, ship it in products.

Under the name is a familiar shape: a mixture-of-experts model with 117 billion parameters in total, of which only about 5 billion are active per token. That is what lets a model with OpenAI on the badge answer quickly and cost $0.60 per million tokens. The honest limitation is the context window: at 128k tokens — roughly 95,000 words — it is the smallest on UsingOpen, fine for most conversations and code files, but not the place to paste an entire repository.

It supports tool use and zero-data-retention terms, and it has become a favourite for agent work: steady behaviour, visible reasoning, and an ecosystem of tooling that grew up around it after its August 2025 release. On UsingOpen it is the natural pick when you want OpenAI-style reasoning in an open model you could, if you wanted, go and run yourself.

What can you do with GPT-OSS 120B?

Where GPT-OSS 120B has found its audience:

  • Step-by-step problem solving — maths, logic, planning — with reasoning you can inspect.
  • Agent workflows that call tools across many turns, where predictable behaviour matters.
  • Coding help on files and modules that fit comfortably in 128k tokens, roughly 95,000 words.
  • Self-hosting projects: Apache-2.0 weights that run on a single high-memory GPU node.
  • Fine-tuning experiments — permissive licence, well-understood behaviour, an active community.
  • Comparing open reasoning models against the closed habit: same prompt, watch the styles differ.

How does GPT-OSS 120B compare?

Its most interesting rival on the site is Meta's Muse Glimmer 30B — another compact open agent, at $1.50 per million tokens with a distillation pedigree and picture reading. Against Z.ai's GLM 5.3 Flash ($0.25), the trade is context and price: GLM's million-token window and image input versus GPT-OSS's OpenAI heritage and Apache-2.0 simplicity.

Against the reasoners, NVIDIA's Nemotron 3 Ultra ($2.40) thinks harder and orchestrates other models, at four times the price. GPT-OSS 120B's pitch is the sensible default: a reasonable price, inspectable reasoning, and weights you can walk away with.

Is GPT-OSS 120B right for you?

GPT-OSS 120B is the right pick when you value inspectable reasoning and an uncomplicated licence: step-by-step working you can check, Apache-2.0 weights you can host or fine-tune without a conversation with legal, and steady behaviour across long tool chains. It has become a default for people building agents who want fewer surprises, and the ecosystem of examples and tooling that grew up around it after its 2025 release is itself a feature — when something misbehaves, someone has usually already written down why.

Its limits are equally clear. The 128k window is the smallest on the site — repository-scale dumps belong to the million-token models — and there is no image input, so visual work goes elsewhere. If your budget is tight, GLM 5.3 Flash undercuts it while adding sight; if you want deeper reasoning per question, Nemotron 3 Ultra costs four times as much and earns it.

The shorthand: pick GPT-OSS when the reasoning needs to be visible, the licence needs to be boring, and the job fits in a normal-sized window. That description covers a surprising amount of serious work.

Frequently asked questions

Is GPT-OSS 120B really made by OpenAI?

Yes. It is OpenAI's open-weight release — its first since the GPT-2 era — announced in August 2025 and shipped under Apache 2.0, with both the weights and the tooling published openly. The release was a genuine event in open AI: the company that defined closed frontier models shipping weights again, with reasoning and tool support intact.

What is the context window of GPT-OSS 120B?

128k tokens — roughly 95,000 words. That is the smallest window on UsingOpen: ample for conversations, documents and individual code files, but not for whole-repository dumps the way the million-token models manage. In practice 128k covers the great majority of conversations, single documents and individual code files — the window only becomes the story when you paste a whole repository.

Is GPT-OSS 120B a reasoning model?

Yes. It is built to reason step by step, and its thinking is visible — UsingOpen shows the reasoning alongside the answer so you can follow the working rather than just the result. The reasoning is adjustable — you can ask for more or less deliberation — which lets you trade latency for care on a question-by-question basis.

Does GPT-OSS 120B support image input?

No. It is a text-and-code model. If you want sight at a similar price, GLM 5.3 Flash and Gemma 4 26B both read pictures and cost less per million tokens.

What does OSS mean in GPT-OSS?

Open-source software — the name marks it as OpenAI's open-weights line. The model is released under Apache 2.0, one of the most permissive licences in the field.

When was GPT-OSS 120B released?

August 2025. It was OpenAI's return to open weights after years of closed frontier models, and it arrived with both a 120B model and a smaller sibling for smaller hardware. It has now been in the wild long enough to accumulate a body of community knowledge: integrations, fine-tunes and honest failure modes are all documented.

Is GPT-OSS 120B free to chat with on UsingOpen?

Yes. Pick it in the model picker and start chatting — no account, no payment, with history saved only on your device. For people deciding whether open reasoning models are ready for their work, this is the cheapest possible due diligence.

Related models on UsingOpen

If GPT-OSS 120B is not quite the right shape, these neighbours on UsingOpen are worth a look:

  • Muse Glimmer 30B by Meta: the compact agent from Meta Superintelligence Labs, distilled to run on ordinary hardware.
  • GLM 5.3 Flash by Z.ai: the fast default — 18B of 320B parameters awake per token, a million-token window and picture input.
  • Nemotron 3 Ultra by NVIDIA: the largest open NVIDIA model — 550B total, 55B awake, built to reason and orchestrate.

Head-to-head comparisons

See how GPT-OSS 120B lines up with the models people most often weigh it against:

Model availability changes as upstream capacity comes and goes — the picker on the chat page shows what is ready right now.

Chat with GPT-OSS 120B free on UsingOpen — no account, no sign-up

← All open-weight models · Compare models · Free AI chat · FAQ