If you follow AI, you have probably had this conversation. "Are you using Cursor or Claude?" "Isn't Claude inside Cursor?" "What about Antigravity, is that Gemini?" "And what is Astra, again?"
The confusion is not your fault. These names look like competitors, but they are not all the same kind of thing. A model, an app, a coding agent and a research prototype can all be called "AI", yet they sit on different layers. Once you see the layers, the whole landscape clicks into place.
This is the first of a short series. Here we build the map. The other articles zoom in on each family, and one is dedicated to the question everyone asks: what is the difference between paying for an API and paying for a subscription?
Everything here was checked on 30 September 2026. AI products change names, limits and prices every few weeks, so treat the links as the source of truth and the numbers as a snapshot.
Layer 1: the model (the brain)
A model is the trained neural network itself. It takes text, images or audio in, and produces text (or images, or audio) out. It has no buttons, no chat window and no memory of you. It is a file of numbers that somebody trained for months on an enormous amount of compute.
Examples: Claude Opus, Gemini Flash, GPT-5.6, Muse Spark.
Models have versions (Sonnet 5.5, Gemini 3.8 Flash) and tiers (a small fast one, a balanced one, a large slow one). Most of the naming confusion lives here, and we will decode each family's names in its own article.
Layer 2: the product (the place you talk to it)
A product wraps a model in an interface: a chat window, a voice mode, a phone app. You do not pick a file of weights, you open an app.
Examples: the Claude app, ChatGPT, the Gemini app, Meta AI.
The product adds things the raw model does not have: conversation history, file uploads, web search, memory, image generation, and safety systems. When people say "I asked ChatGPT", they mean the product. When a developer says "I called gpt-5.6", they mean the model.
Layer 3: the harness (the model with hands)
This is the layer that changed the game in the last two years. A harness (you will also hear agent or coding agent) gives a model tools: it can read your files, run commands in a terminal, edit code, browse the web, call other programs, and then look at the result and decide what to do next.
The key idea is a loop:
- The model reads the situation and decides on an action ("open this file").
- The harness carries out the action.
- The result goes back to the model.
- Repeat until the task is done.
That loop is what people mean by agent mode. A chatbot answers. An agent acts.
Examples: Claude Code, OpenAI Codex, Gemini CLI, Cursor's agent, Google Antigravity.
Notice that some of these are made by the model maker (Claude Code by Anthropic, Codex by OpenAI, Gemini CLI and Antigravity by Google), and one is made by a separate company that uses other people's models (Cursor).
Layer 4: the surface (where the harness lives)
The same harness can appear in several surfaces:
| Surface | What it feels like |
|---|---|
| Terminal / CLI | You type in a command line. Scriptable, works over SSH and in CI. |
| IDE / editor | The agent lives in your code editor, with diffs and inline edits. |
| Desktop app | A standalone window for managing several agents at once. |
| Web / mobile | Kick off a task from a browser or phone and let it run in the cloud. |
For example, Claude Code is available in a terminal, in editor extensions, in a desktop app and on the web, all driven by the same engine.
Where each name sits
| Name | Layer | Made by |
|---|---|---|
| Claude Haiku, Sonnet, Opus, Fable | models | Anthropic |
| Claude (the app), Claude Code | product, harness | Anthropic |
| GPT-5.6 (Luna, Terra, Sol) | models | OpenAI |
| ChatGPT, Codex | product, harness | OpenAI |
| Gemini (Pro, Flash, Flash-Lite) | models | |
| Gemini app, Gemini CLI, Antigravity | product, harnesses | |
| Project Astra | research prototype, feeding Gemini Live | Google DeepMind |
| Muse Spark | model (and the Meta AI product) | Meta |
| Cursor | editor and agent that uses many models | Anysphere (the Cursor company) |
The most useful sentence in this whole article: Cursor is to models what a web browser is to websites. It is a place to use models from several companies, plus a few of its own.
The other map: how you pay
There are two completely different ways to pay for AI, and mixing them up is the most expensive mistake beginners make.
- A subscription is a flat monthly fee for a product (the Claude app, ChatGPT, Google AI, Cursor). You get a usage allowance, reset on a schedule, inside that company's apps.
- An API is metered access to the model itself, billed per token (a token is roughly three quarters of a word). You write the code, and you pay for exactly what you send and receive.
The same company usually offers both, and they are separate accounts with separate bills. A ChatGPT subscription, for example, does not include API credits. We give this its own article, with a decision guide, because it matters so much.
What you can do with all this
- Learn and write faster: explain a concept five ways, draft, summarise, translate.
- Build software without typing every line: describe a feature, let an agent write it, run the tests and fix failures.
- Automate the boring parts: a script that reviews every pull request, triages support email, or summarises logs each morning.
- Build your own AI features: call an API from your app to add search, chat, extraction or classification.
- Explore a field: read long documents, compare sources, analyse a dataset.
The rest of the series
- Claude explained: Haiku, Sonnet, Opus and Fable
- OpenAI explained: GPT, ChatGPT and Codex
- Google's AI stack: Gemini, Antigravity and Project Astra
- Cursor explained: the AI editor that uses everyone's models
- Meta's Muse Spark and the Muse family
- API versus subscription: how you actually pay for AI
- CLIs and agent mode: the terminal agents compared
If you want to understand what is happening inside these models, our free Deep Learning course builds up from a single neuron to backpropagation, optimizers and regularization, with simulators to experiment in.