Yes, ChatGPT is an LLM, or more precisely, it runs on one. A large language model (LLM) is a neural network trained to predict the next piece of text, and every reply ChatGPT writes comes out of one. ChatGPT itself is the product wrapped around that model, from the chat window to the tools it can call, and the model inside it has changed many times since launch.
The answer used to be tidier. When OpenAI launched ChatGPT on November 30, 2022, the announcement opened with "We've trained a model called ChatGPT." The name belonged to the model. Today it belongs to the app, and the LLM answering you depends on which plan you pay for.
So the honest answer is yes, with an asterisk. The asterisk has grown since 2022, and it now needs its own release notes page.
Is ChatGPT a neural network?
Yes. The LLM inside ChatGPT is a neural network, a program built from layers of simple numeric units whose connection strengths, called weights or parameters, are learned from data.
The specific design is a transformer, the architecture Google researchers introduced in 2017. A transformer reads text as tokens, meaning chunks of roughly three quarters of a word, and uses a mechanism called attention to decide how much each earlier token matters to the one it is about to produce. GPT stands for generative pre-trained transformer, and every word in that acronym is literal.
Generation is a loop. The model scores every possible next token and one gets sampled. That token joins the text, and the loop runs again. A 500-word answer takes several hundred trips round that loop, which is why replies can stream onto the screen a few words at a time.
Everything that reads like reasoning in a ChatGPT reply is built from that single move, repeated at scale. Our explainer on how a large language model is made, step by step walks through tokenization and pretraining. For the machinery below the API, LLM Systems Engineering covers embeddings and the transformer itself in working detail.
How LLM training turned a text predictor into ChatGPT
A freshly pretrained LLM is a brilliant autocomplete with no manners. Ask it a question and it may reply with three more questions, because that is what text on the internet often does next.
OpenAI's fix, published as InstructGPT in 2022, added two stages after pretraining. Supervised fine-tuning trained the model on example answers written by people. Reinforcement learning from human feedback (RLHF) came second: raters ranked competing answers, and a separate reward model learned to predict their rankings. The LLM was then tuned to score well against that reward model.
The payoff was large. In the InstructGPT paper, human evaluators preferred outputs from a 1.3 billion parameter tuned model over those of the original 175 billion parameter GPT-3, despite the tuned model having 100 times fewer parameters. Post-training, meaning everything done to a model after pretraining, turned out to matter more than raw size for whether people found the model useful.
ChatGPT was built the same way. OpenAI's launch post describes fine-tuning a GPT-3.5 series model with RLHF, using conversations in which human trainers played both the user and the assistant. The special tokens that mark whose turn it is in a conversation get their meaning in this stage too, a detail covered in how special chat tokens are trained.
The same post admitted the model was often excessively verbose, because trainers tended to prefer longer answers that looked comprehensive. Anyone who has asked ChatGPT a yes or no question and received five paragraphs has met that preference in person.
Which LLM does ChatGPT use in 2026?
Several, and the answer changes with your plan and with the month. As of October 2026, the Chat tab runs on OpenAI's GPT-6 family.
According to the ChatGPT release notes, the rollout that began on October 7 assigns models by plan:
| Plan or option | LLM answering in the Chat tab |
|---|---|
| Free and Go | GPT-6 Luna |
| Plus, Pro, Business and Enterprise | GPT-6 Sol |
| Pro reasoning option | GPT-6 Astra |
Two people typing the identical question into ChatGPT on different plans are talking to two different LLMs. The Work and Codex modes run their own lineup on top of that.
The lineup also turns over fast. GPT-4o and the original GPT-5 left ChatGPT on February 13, 2026, and GPT-5.1 followed on March 11. The default model has changed twice since May alone. Any article from 2024 that confidently names the model behind ChatGPT is now a historical document.
Routing has changed too. In September, OpenAI retired automatic switching from Instant to Thinking for Plus and Pro users, though the release notes say ChatGPT can still switch models automatically for safety purposes. Even the LLM handling a single conversation is a decision the product makes on your behalf.
What ChatGPT adds around the LLM
The LLM produces text. Most of what you experience in ChatGPT beyond that is ordinary software built around the model.
| Layer | What it does | How the LLM is involved |
|---|---|---|
| System prompt | Hidden instructions sent ahead of your message | The model reads them as text |
| Memory | Saved details and past chats about you | Stored outside the model, supplied as context |
| Web search | Fetches pages for questions about current events | The model summarises what comes back |
| Tools and apps | Runs code and reaches connected services | The model decides when to call them |
| Safety checks | Moderation and model switching on sensitive topics | Runs alongside the model |
The weights never change while you chat. When ChatGPT remembers your dog's name, a piece of software saved a note and slipped it back into the prompt. The model's affection for your dog remains purely statistical.
The same split explains why an LLM's web search answers can be inaccurate. The search step can retrieve a weak or outdated page, and the model can misread a good page while summarising it. Both failures arrive in the same confident tone, so open the cited source before you repeat the claim.
Is ChatGPT an AI agent?
Partly. An LLM agent is an LLM running in a loop with tools. The model picks an action and software carries it out. The result comes back, and the model decides what to do next, until the task is done.
A single reply in the Chat tab works more like a function call. You ask and it answers, and nothing further happens until you type again. Even when it searches the web mid-answer, the loop is short and you are steering it.
ChatGPT now contains agents as well. The release notes describe dots, always-on agents rolling out to some paid plans, which run on GPT-6 Astra and keep working between conversations on their own cloud computer. Work and Codex modes run multi-step tasks with far less supervision than a chat reply.
Every one of those agents has an LLM at its centre doing the deciding. The engineering effort goes into the tools and the stopping rules around it, which is the ground Agentic AI Engineering covers for teams building their own.
Why the difference matters when you build
If you build on OpenAI's API, you get the LLM without ChatGPT's product layer. None of the memory or hidden instructions come along, and each request is stateless, so your code resends the whole conversation every time. Our guide to what an AI API actually returns covers that request cycle in detail.
This catches teams moving from prototype to product. A prompt tuned in the chat window often behaves differently through the API, because the chat window was adding instructions you never saw.
A short checklist for anyone putting the answer to use:
- Name the model in your notes and specs. "ChatGPT" covers several LLMs that change by plan and by month.
- Test prompts against the exact API model you will ship on, and keep the chat window for exploration.
- Treat web search answers as leads, and open the cited page before relying on a claim.
- List the memory and tools a workflow quietly depends on, then rebuild those parts deliberately when you move to the API.
Next time someone asks whether ChatGPT is an LLM, you can say yes in one word. Whether you then spend ten minutes on the asterisk is between you and your friendships.