A minimal glossary and a comparison of LLMs

The terms you should know

Let’s clear up some terms that are often used as synonyms, wrongly. You’ll find these and many other terms in the Glossary, with more complete definitions. We’re covering them here too because getting them straight from the start really matters.

AI (Artificial Intelligence): the branch of computer science dealing with systems that simulate cognitive abilities (I know, it’s imprecise and reductive, don’t hold it against me).

LLM (Large Language Model): large language models, pre-trained on enormous amounts of text, capable of generating, completing, transforming, and analyzing language.

GPT (Generative Pre-trained Transformer): the term has two uses. Strictly speaking, it refers to OpenAI’s models (GPT-3, GPT-4, and later ones). More loosely, it has become almost a synonym for “LLM,” because it’s by far the most common type of model, the one based on the transformer architecture.

ChatGPT: the interface for the LLM created by OpenAI (an American company); it was the first to amaze the general public in late 2022 (see AI Timeline). ChatGPT and GPT often end up being used as synonyms. If you’re confused about this, you’re right to be — it is confusing, but that’s how it is.

Token: the smallest unit of text an LLM processes. It doesn’t correspond to a word: it can be a whole word, part of a word, a space, or a punctuation mark. On average, a token is about 3/4 of a word. This matters because tokens determine how much you can write in a chat and how much you pay.

Prompt: the text you write to the LLM, your request. It can be a short question or a long, detailed instruction. The quality of the prompt directly affects the quality of the response.

Chat: a single conversation with the LLM, with a beginning and an end. Each chat has its own temporary memory: the model remembers what you’ve written since the start of that chat, but not previous chats (except for specific memories).

Model: the trained “brain” that generates the responses. It shouldn’t be confused with the interface: ChatGPT is the interface, GPT-5.6 Sol is the model running behind it.

Interface: the site or app you talk to the model through. ChatGPT, Claude, Gemini, Le Chat are interfaces; the underlying models are a different thing, and often you can choose which one to use.

Hallucination: when the LLM generates false information but presents it confidently, as if it were true. It happens because the model predicts plausible text, it doesn’t verify facts.

Context (context window): the maximum amount of text an LLM can take into account in a single conversation. The larger it is, the more information the model keeps in mind, but it’s not infinite.

Deep Research: a mode in which the LLM autonomously runs dozens of web searches, analyzes them, and produces a long, well-documented report. Slower and more expensive than a normal response, but much more thorough.

Credits: the usage budget included in a subscription. They get used up with every message, faster when you use powerful models or advanced modes like Deep Research.

Subscription: the paid plan (usually monthly) that gives access to more powerful models and more credits. Free plans exist but have significant limits.

Sycophancy: the model’s tendency to agree with us, praise us, and avoid contradicting us, even when it should. An English term similar to “compiacenza” but more specific.

Epistemia: the illusion of possessing genuine knowledge because the AI provides eloquent, convincing answers. Mistaking form for substance.

Open weights: when a model can be downloaded and run on your own computer (with adequate hardware), without depending on the provider.

Frontier model: the most advanced and powerful models available at a given moment. The meaning of the term shifts every few months because progress is extremely fast.

Comparing the models

Frontier models

Other models