Is ChatGPT Open Source? What Is Open and What Is Not (2026)
Is ChatGPT open source? No. ChatGPT is a closed, proprietary service, and OpenAI does not publish its code or its model weights. OpenAI has released separate open-weight models called gpt-oss, which anyone can download and run themselves, but those models are not part of ChatGPT and do not appear inside it.
If you are wondering whether ChatGPT is open source, the short version is that the chatbot is not, but OpenAI does give away some AI models you can run on your own computer. The two ideas get mixed up all the time, so this guide separates them. You will see what is open, what is closed, what open-weight means and how to try an open model safely.
In this guide
Is ChatGPT open source?
No. ChatGPT, the website and apps you reach at chatgpt.com, is a closed service run by OpenAI. OpenAI does not publish the source code or the trained model weights behind it, so nobody outside the company can inspect, change or host their own copy. You can only use it through OpenAI's apps, website and API.
OpenAI's own help centre confirms the split. In its article on OpenAI open-weight models (gpt-oss) it says: "These models are not served through the OpenAI API and are not available in ChatGPT." So when you open ChatGPT today, you are not talking to an open model.
Why does it matter? Open software lets anyone check how it works. With ChatGPT you cannot, so you rely on OpenAI's published policies for how your chats are handled. Our guide to whether ChatGPT is safe covers that side in plain English.
What is the difference between open source and open weight?
Open source normally means the code is public and you are allowed to use, change and share it. Open weight is narrower. The weights are the numbers a model learns during training. When a company publishes them, you can download the model and run it yourself, even if the training data and the full training process stay private.
OpenAI itself avoids calling gpt-oss fully open source. Its help article explains that it uses the terms open models and open-weight, meaning the trained weights are public under the Apache 2.0 licence and the gpt-oss usage policy. It also notes that some surrounding tooling may stay proprietary.
Open source: the code is public and reusable.
Open weight: the trained model files are public, but not necessarily everything used to make them.
Closed, like ChatGPT: you can use the service, but you cannot download or inspect the model.
What has OpenAI actually opened?
OpenAI has opened a few separate things, and none of them is ChatGPT itself. This table sums up the position as we read it on 24 October 2026.
Product | Open or closed | What you can do |
ChatGPT website and apps | Closed | Use it through OpenAI's own apps and site |
Models in the OpenAI API | Closed | Call them through the API and pay by usage |
gpt-oss-120b and gpt-oss-20b | Open-weight, Apache 2.0 | Download, run and fine-tune on your own hardware |
gpt-oss-safeguard | Open-weight, Apache 2.0 | Download and run safety classification models |
Codex CLI, SDK, skills and plugins | Open source on GitHub | Read the code and report issues |
Codex IDE extension and Codex cloud | Not open source | Use them as products |
gpt-oss: the open-weight models
OpenAI released gpt-oss-120b and gpt-oss-20b on 5 August 2025 under the Apache 2.0 licence. In its announcement it described them as its first open-weight language models since GPT-2. The larger model has about 117 billion parameters and fits on a single 80 GB GPU. The smaller one has about 21 billion and needs around 16 GB of memory, so a good home computer can run it. Both are text-only and support a context length of up to 128,000 tokens.
On the benchmark table on OpenAI's open models page, gpt-oss-120b scored 90.0 on the MMLU knowledge test against 93.0 for o4-mini, and 80.1 on GPQA Diamond against 81.4. That puts it close to a smaller OpenAI reasoning model. These are OpenAI's own figures, so treat them as a guide and not as proof of how it behaves in everyday use.
gpt-oss-safeguard
On 29 October 2025 OpenAI added gpt-oss-safeguard, a pair of open-weight models built on gpt-oss for sorting content against a written safety policy. They are meant for developers and trust and safety teams, not for everyday chat.
Codex and the other open parts
OpenAI's open-source documentation for Codex lists the Codex CLI, the SDK, reusable skills and plugins as developed in the open on GitHub. It says the IDE extension and Codex cloud are not open source. The main repository takes bug reports but does not accept outside code contributions. Our guide to ChatGPT Codex explains what the product does.
Can you run an open OpenAI model yourself?
Yes, if your computer is powerful enough. OpenAI's Cookbook guide to running gpt-oss with Ollama uses a free tool called Ollama to run the model on a normal PC or Mac. The guide recommends at least 16 GB of VRAM or unified memory for gpt-oss-20b and at least 60 GB for gpt-oss-120b. You can lean on the processor if you are short on memory, but it will be slower.
The steps, from OpenAI's guide
Check your memory. About 16 GB of VRAM or unified memory is the guide's recommendation for the 20b model.
Install Ollama from ollama.com/download.
Download the model. Open a terminal and type: ollama pull gpt-oss:20b
Start a chat. Type: ollama run gpt-oss:20b
Ask your question. Once the download has finished, the model runs on your own machine, and you can chat with it offline.
Illustrative example, not a test we ran: a small accountancy firm wants a model to summarise client letters without sending the text to an outside company. Running gpt-oss-20b on a company computer keeps the text on that machine. The firm would still need a person to check every summary, because open models make mistakes just as hosted ones do.
A prompt you can paste once the model is running: "Summarise the text below in five bullet points for a non-expert. List anything you are unsure about at the end."
What are the limits and privacy points?
It is not ChatGPT. There is no ChatGPT app, memory, voice or image tools. OpenAI describes gpt-oss as text-only reasoning models.
Free weights are not free running. OpenAI says you are responsible for the compute, storage or hosting costs.
There is no support desk. OpenAI says it does not provide hands-on help or debugging for self-hosted or third-party-hosted setups.
Your data stays with you, if you host it yourself. OpenAI says it does not receive the data you send to a self-hosted model unless you share it or use one of its managed hosting partners. If you use another company's hosting, that company's privacy terms apply.
A usage policy still applies. The Apache 2.0 licence allows commercial use, subject to the gpt-oss usage policy.
For how ChatGPT itself treats your conversations, read Does ChatGPT save your conversations?.
Should you use ChatGPT or an open model?
For most people ChatGPT is simpler. There is no hardware to buy, and it works on every device. An open model makes sense if you need text to stay on your own systems, want to fine-tune a model, or simply enjoy tinkering.
If you are new to all of this, start with What Is ChatGPT?. Developers weighing up costs can read ChatGPT API pricing explained, and anyone comparing chatbots may like our ChatGPT vs DeepSeek comparison. For background on how these systems work, see large language models explained.
Frequently asked questions
Is OpenAI an open source company?
Not in the sense most people mean. OpenAI keeps the models behind ChatGPT and its API closed. It does publish some open-weight models, called gpt-oss, and some tools such as the Codex CLI on GitHub.
Are the GPT models behind ChatGPT open source?
No. OpenAI's help centre says its open-weight gpt-oss models are not served through the OpenAI API and are not available in ChatGPT, so the models you chat with there remain closed.
Can I download ChatGPT and run it offline?
No. ChatGPT needs OpenAI's servers, and the desktop and mobile apps are only ways to reach it. You can download and run gpt-oss offline instead, but that is a different, open-weight model.
Is gpt-oss free to use?
The weights are free to download under the Apache 2.0 licence. OpenAI points out that you pay for the compute, storage or hosting you use to run them, so the model is free but the running costs are yours.
Can I use gpt-oss in a business?
OpenAI's help article says the Apache 2.0 licence allows commercial use, subject to the gpt-oss usage policy. Read the policy and your own data rules before building anything customer-facing.
Is gpt-oss as good as ChatGPT?
OpenAI's published benchmarks put gpt-oss-120b close to its o4-mini model on several tests. Those are OpenAI's own figures, and a benchmark score does not tell you how a model feels in everyday chat or whether it matches the latest ChatGPT models.
Key Takeaways
ChatGPT is closed. OpenAI does not publish its code or model weights.
OpenAI's open-weight models are gpt-oss-120b, gpt-oss-20b and the gpt-oss-safeguard pair, under the Apache 2.0 licence.
Open weight is not the same as open source, and OpenAI itself uses the narrower term.
You can run gpt-oss-20b at home with Ollama if you have about 16 GB of VRAM or unified memory.
Self-hosting keeps your text on your machine but leaves the cost, setup and checking to you.
Conclusion
ChatGPT is not open source, but OpenAI does publish open-weight models called gpt-oss that anyone can download and run under the Apache 2.0 licence. The practical question is whether you need that control. Most people are better served by the ChatGPT app, while developers and privacy-minded teams may want a local model. Open models are moving quickly, so recheck OpenAI's help article before you rely on any detail here. Explore more plain-English ChatGPT guides on whatischatgpt.co.uk.
References
Checked on 24 October 2026. We read OpenAI's help article, the gpt-oss announcement, the open models page, the Cookbook guide and the Codex open-source page. We did not run these models ourselves for this guide, so the commands come from OpenAI's documentation. We are independent of OpenAI, and its products and policies can change without notice.
