Is ChatGPT open source?
No. The ChatGPT application and the GPT-5.6-series models behind it are closed — no source, no weights, no training data. What confuses this question is that OpenAI genuinely does publish other models openly: gpt-oss-120b, gpt-oss-20b and gpt-oss-safeguard, under Apache 2.0, downloadable and runnable on your own hardware. Those are not ChatGPT, and they are open-weight rather than open-source.
Both halves are true, which is why everyone gets this wrong.
The closed half. Nothing about ChatGPT is published as source. You cannot download the weights of GPT-5.6 Luna or Sol, inspect the application, self-host the product or fork it. Access is through OpenAI's apps or its paid API, under its terms — which, per its own Pro article, prohibit reselling access or using ChatGPT to power third-party services. That is the ordinary shape of a commercial product, and it is the answer to the question as asked.
The open half, which is genuinely open. OpenAI publishes a family called gpt-oss — reasoning models at 120B and 20B parameters, plus gpt-oss-safeguard for custom safety policies — described as "designed to run locally on desktops, laptops, and in data centers". The licence is Apache 2.0, which OpenAI describes as permissive, without copyleft restrictions or patent risk, for experimentation, customisation or commercial deployment. There are official guides for running them under Ollama, LM Studio, vLLM and Transformers.
The distinction that resolves it. gpt-oss is open-weight, not open-source in the sense a software project is. You get the trained parameters and permission to use them; you do not get the training data, the training code, or the process that produced them. That is a meaningful amount of freedom — you can run it offline, fine-tune it, and inspect its full chain-of-thought — and it is not the same as a repository you could rebuild from scratch.
So the fully accurate sentence is: OpenAI ships open-weight models; ChatGPT is not one of them. Anyone who tells you ChatGPT is open source has collapsed those two facts into one, and anyone who tells you OpenAI publishes nothing openly has missed the second.
What you can actually do with the open ones.
gpt-oss is a real answer to that, and an unusually good one given who published it. The 20B model is sized for a decent laptop or desktop; the 120B expects serious hardware or a data centre. OpenAI's own documentation points at Ollama, LM Studio, vLLM and Transformers, so this is a supported path rather than a community reverse-engineering effort.
What you get that ChatGPT will never give you: it runs with no network, nothing leaves your machine, there is no allowance to exhaust, and you can fine-tune it on your own material. What you give up is everything built around the model — the apps, the tools, the memory, the file handling, the agentic surfaces.
OpenAI publishes benchmark comparisons on its own open-models page, putting gpt-oss-120b and gpt-oss-20b alongside OpenAI o3 and o4-mini across MMLU, GPQA Diamond, Humanity's Last Exam and competition maths. We reproduce no numbers from it here, for a reason we apply everywhere: those are vendor-reported figures for a vendor's own models, and we did not run them.
The structural point stands without any number: gpt-oss is a separate, smaller, openly licensed family, and it is not the model answering you in ChatGPT. If you want a like-for-like sense of the trade, run it yourself on your own task — which is, conveniently, exactly what an open-weight licence lets you do.
Three things collide. The company is called OpenAI. It really does publish open-weight models. And "open source" is used loosely enough in AI coverage to mean anything from a downloadable checkpoint to a fully reproducible training pipeline.
We hit the same pattern in the other direction on Claude Code, where the AI Overview asserted the tool was open source and the repository's own licence said otherwise. The lesson generalises: for licence questions, the licence file and the vendor's own licensing page are the only sources worth trusting, and both take under a minute to check.
Models you can actually download.
All three are open-weight and runnable on your own hardware. We have not benchmarked any of them — quality tracks whatever your machine can host, so test on your own task before committing to one.
Frequently asked.
Quick follow-ups people search after this question.