gpt-oss playground
A free web playground published by OpenAI where developers can chat with the open-weight models gpt-oss-120b and gpt-oss-20b, adjust reasoning effort, watch the reasoning unfold and copy commands to run the same models locally.
What is gpt-oss playground?
gpt-oss playground is the demonstration site OpenAI put online for its two open-weight language models, gpt-oss-120b and gpt-oss-20b. Its own description is deliberately narrow: a demo of those models, for developers. The publisher's launch announcement points readers to it directly, which is what identifies the site as first-party rather than one of the many unofficial mirrors that share the model's name.
It is a single page and nothing more. There is no pricing page, no about page, no documentation, no blog and no login wall: /terms, /privacy, /about, /pricing and /docs all return a 404, and the domain carries neither a robots.txt nor a sitemap. Everything the site does happens in one chat surface, built with ChatKit and backed by an api.gpt-oss.com endpoint.
On first load a short dialog asks a single question — whether reasoning should be visible — and warns that it may be verbose, speculative or off. Past that, the visitor picks a model and starts typing. The two models are presented by contrast: gpt-oss-120b is designed for large-scale infrastructure, gpt-oss-20b is optimized for running on-device. A three-step control sets the reasoning effort to High, Medium or Low, and the chain of reasoning can be shown or hidden at will. Four starter prompts are offered, ranging from a general-knowledge question to a short Python exercise and two classic reasoning traps.
The most useful part for a developer sits below the chat. A hosting provider panel turns the playground into a launchpad: for each model it prints the exact command to run it elsewhere — through Ollama, LM Studio, or Hugging Face Transformers — with links to each tool. Signing in with a Hugging Face account is offered but never required.
According to the publisher's announcement, both models are released under the Apache 2.0 licence, gpt-oss-120b approaches o4-mini on core reasoning benchmarks while fitting on a single 80 GB GPU, and gpt-oss-20b performs comparably to o3-mini within 16 GB of memory. Neither is served through the OpenAI API, which is precisely why a public playground exists at all.
What it does
- Chat with gpt-oss-120b or gpt-oss-20b directly in the browser
- Switch between the two open-weight models from the model picker
- Set the reasoning effort to High, Medium or Low
- Show or hide the model's chain of reasoning
- Copy the exact command needed to run either model locally
- Sign in with a Hugging Face account, entirely optionally
- Switch the interface between light and dark themes
When to use gpt-oss playground / When not to
A quick filter to help you decide if gpt-oss playground is the right fit.
When to use gpt-oss playground
- Developers who want to try gpt-oss-120b and gpt-oss-20b without provisioning a GPU
- Engineering teams deciding which of the two model sizes to deploy
- ML and MLOps practitioners who need the exact Ollama, LM Studio or Hugging Face command to run the models locally
- Researchers and AI safety specialists who want to inspect a visible chain of reasoning
- Lecturers and students who need a live demonstration of an open-weight reasoning model
When not to use gpt-oss playground
- Teams needing a production endpoint: the site calls itself a demo and makes no service commitment
- Anyone handling confidential, personal or regulated data, since no privacy policy or terms are published
- Users who need to upload documents or images, as attachments are disabled in the composer
- Developers looking for a documented public API to call from their own code
- Buyers who expect support, an account, billing or any contact channel
How to use gpt-oss playground
A typical end-to-end flow, from setup to results.
- Open gpt-oss.com in a browser; no sign-up and no payment details are required
- On the welcome dialog, choose Continue with visible reasoning or Don't show reasoning
- Pick a model: gpt-oss-120b for large-scale infrastructure, or gpt-oss-20b for on-device use
- Set the reasoning level to High, Medium or Low
- Type a request in the Ask anything composer, or start from one of the four suggested prompts
- Read the answer and, when visible reasoning is on, the reasoning that produced it
- Change model, reasoning level or reasoning display at any point in the session
- Open the Hosting provider panel to see how to run the same model yourself
- Copy the command for your tool: ollama pull gpt-oss:20b then ollama run gpt-oss:20b, or lms get openai/gpt-oss-20b, or pip install transformers torch then transformers serve
- Optionally sign in with Hugging Face; the playground works without it
Pros & Cons
Pros
- Immediate access to both models with no account, no card and no installation
- Direct comparison of the 120b and 20b variants before committing to either
- Visible chain of reasoning, which few public playgrounds expose
- Reasoning effort adjustable across three levels
- A clear exit path to self-hosting, with the exact command for three different tools
- Models released under the Apache 2.0 licence according to the publisher's announcement
- First-party site operated by OpenAI rather than an unofficial mirror
Cons
- No privacy policy, terms of service or legal notice anywhere on the domain
- No email address, contact page or support channel of any kind
- No documented public API; the gpt-oss models are not served through the OpenAI API
- No file uploads, as attachments are disabled in the composer
- No stated usage limits, quotas or availability commitments
- Explicitly labelled a demo, so it carries no service guarantee
- English-only interface with no language selector
Pricing & Plans
The playground is free to use. gpt-oss.com publishes no pricing page — the /pricing path returns a 404 — and the application bundle contains no reference to a plan, a price, a credit, a quota or a paid account. No payment path exists, and no account is required to start a conversation. There is consequently no entry-level paid price to quote and no currency to state. Any cost attached to these models arises elsewhere: running them on your own hardware, or through a third-party host such as Hugging Face, Ollama or LM Studio, none of which is priced by this site.
Data, GDPR & hosting
A consolidated view of how gpt-oss playground handles your data.
GDPR overview
There is no GDPR information of any kind on gpt-oss.com. The terms GDPR, DPA, data processing agreement and sub-processor do not appear once in the collected content, which covers the homepage, the archived HTML and the complete JavaScript bundle carrying the whole single-page application. No privacy policy, terms, legal notice or cookie banner is published: those paths all return a 404. No representative under Article 27 is designated, no data protection officer is named, and the domain exposes no email address at all. This is an absence of published information rather than a statement of non-compliance, and it should be read that way: the site is a demonstration surface, and any GDPR assessment would have to rest on documents this domain does not carry.
Who owns the data?
gpt-oss.com publishes no terms and conditions and no privacy policy, so ownership of prompts, conversations and model outputs is simply not addressed anywhere on the site. Requesting /terms, /privacy or /legal returns a plain 404, and the words privacy, terms and cookie appear nowhere in the page or in its JavaScript bundle. Conversations are handled by an api.gpt-oss.com endpoint through OpenAI's ChatKit component, with no contractual document attached on this domain. The optional Hugging Face sign-in requests the openid, profile, email and inference-api scopes, so a user who chooses to sign in shares profile and email data with that flow.
Reuse rights
Nothing on gpt-oss.com states what the publisher does with what users type, and nothing states what users may do with the answers they receive. There is no licence clause covering model outputs, no statement on whether prompts are used to train or improve models, and no mention of subprocessors or of a data processing agreement. A dedicated search across the homepage, the archived HTML and the full application bundle for opt-out, do not train, disable training and retention returned no match at all. Vercel Web Analytics is loaded on the page, without an accompanying cookie notice. In practice, nothing is contractually settled here, and users should proceed on that basis.
Data retention & training
Hosting summary
No document on gpt-oss.com states where user data is hosted, in which jurisdiction it sits, or under what safeguards it travels. There is no trust or security page, no certification claim and no subprocessor list. What can be observed is technical rather than contractual: the domain resolves to 76.76.21.21, an anycast address served by Vercel and geolocated in the United States; the conversation is handled by an api.gpt-oss.com endpoint; and the chat component is loaded from cdn.platform.openai.com. The domain itself was registered on 16/07/2025 through MarkMonitor, a corporate registrar based in the United States. None of this amounts to a hosting commitment and none of it should be read as one. Anyone with a data residency requirement will find nothing here to satisfy it, and should either run the models themselves — which the site actively encourages — or choose a host that publishes such terms.
Things to keep in mind
Risks and trade-offs to weigh before adopting gpt-oss playground.
- The site calls itself a demo, so it offers no availability guarantee and could change or disappear without notice
- The publisher itself warns that the reasoning shown may be verbose, speculative or simply wrong
- With no privacy policy or terms published, what happens to anything typed into the box is undocumented
- Visible reasoning is persuasive and can make a wrong answer look carefully argued; it is not proof of correctness
- Signing in with Hugging Face shares profile and email data and opens the inference-api scope
- Vercel Web Analytics runs on the page without an accompanying cookie notice
- Treating a free demo as a stable dependency, or mistaking it for the OpenAI API, which does not serve these models
Setup & Integrations
Technical difficulty
Using the playground requires no setup at all: open the page, dismiss one dialog and start typing. No installation, no account, no payment details. The optional Hugging Face sign-in adds a single OAuth step. Running the models yourself is a different matter and assumes comfort with a command line: the site hands over ready-made commands for Ollama, LM Studio and Hugging Face Transformers, but the hardware requirement is real — roughly 16 GB of memory for gpt-oss-20b, and a single 80 GB GPU for gpt-oss-120b, according to the publisher's announcement.
Deployment
Integrations
Supported languages
Behind gpt-oss playground
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Who publishes gpt-oss.com?
Do I need an account to use it?
How much does it cost?
Which models can I try?
Can I see how the model reasons?
Can I run these models on my own machine?
Is there an API?
Can I attach files to a conversation?
Where is the privacy policy?
Is there a mobile app?
Should you pick gpt-oss playground?
gpt-oss playground is exactly what it says it is, and its honesty about that is its main virtue. OpenAI needed somewhere for developers to try two models it does not serve through its own API, and this is that place: one page, two models, a reasoning dial, and the command you need to stop using the page and run the thing yourself. As a way to answer the only question that matters before adopting an open-weight model — is the smaller one good enough for what I need? — it is efficient in a way that a benchmark table is not.
What it is not is a product. There is no privacy policy, no terms of service, no legal notice, no contact address and no support channel of any kind; the eleven usual paths all return a clean 404. That matters less for a throwaway question about the distance to the sun than it does for anything resembling real work, and the sensible line is to keep confidential material out of it entirely. The publisher's own footer says demo only and warns that the reasoning shown may be verbose, speculative or wrong — a warning worth taking literally, because visible reasoning is persuasive even when it is mistaken.
Read as a shop window rather than a workshop, it does its job well. Developers, ML engineers and researchers get a zero-friction look at both models and leave with a working local setup; anyone expecting an SLA, an API or a data processing agreement should look at self-hosting or a commercial provider instead. The models themselves, Apache 2.0 licensed, are where the durable value sits — this site is simply the shortest path to them.
- Choosing a selection results in a full page refresh.
- Opens in a new window.