The model

A model is a computer program that has read a huge amount of text. From that reading, it learned which words tend to follow other words.

When you ask a question, the model writes an answer one small piece at a time. Each piece is its best guess at what comes next.

Think of a new starter at work who has read every book in the library. They know a lot. But they have never worked at your company, and they forget each talk as soon as it ends.

A model is the trained engine inside an AI product. It was built by processing a very large body of text and code, and it learned the patterns in that material.

When you send a request, the model predicts a response piece by piece. It does not look facts up in a database. It generates text that fits what it learned and the request in front of it.

GPT-6.1 Sol and GPT-6 Luna are examples of models in Codex as of October 2026. Model names change often. The ideas in this lesson do not.

A large language model is a transformer network trained by self-supervised next-token prediction over a large corpus of text and code. It is then post-trained with supervised fine-tuning and reinforcement learning to follow instructions, reason and call tools.

At inference time the network maps a token sequence to a probability distribution over the next token. The serving software samples a token, appends it and repeats. Everything the model appears to know at that moment comes from its weights or from the tokens in its input.

The weights do not change while you use the model. A product that seems to remember an earlier session does so by sending stored text back in as input. Lesson 02 covers that.

The model breaks your words into small pieces called tokens. A token is often part of a word.

Then it picks the next token, and the next, until the answer is done. There is a little chance in each pick. So the same question can get a different answer each time.

Some models stop and work through a problem before they answer. This takes longer, but it helps with hard tasks like sums and plans.

Text is split into tokens, which are short pieces of words. Usage limits are counted in tokens, so a long document uses more of your allowance than a short one.

Each answer is generated, not retrieved. The same request can produce different wording, and sometimes different facts. Treat any single answer as a draft.

Reasoning models spend extra effort on a problem before they reply. In Codex you set this as the reasoning effort: low for routine tasks, higher for complex ones. Higher effort takes longer and uses more allowance.

Tokenisation uses a learned subword vocabulary, so cost and context limits are measured in tokens, not words. Numbers, code and text in languages other than English often take more tokens per word.

Sampling at a temperature above zero makes outputs vary from run to run, and serving can add variation of its own. Reproducibility has to come from the system around the model: fixed inputs, saved outputs, and code that the model writes and you can run again.

Reasoning models generate intermediate tokens before the final answer. Codex exposes that budget as reasoning effort, from Light to Ultra as of October 2026. More effort tends to help on multi-step problems, with diminishing returns and more latency.