Models, products, and platforms

“We should use AI for X” can mean 3 different things. It can mean the model that does the generating, the product wrapped around that model, or the platform that serves the model to whoever is building their own. The cost, the control and where the data goes are different at each of the 3, and the people in the room are often not talking about the same one.

The 3 layers

A model is the part that does the generating. GPT, Claude, Gemini, Llama, Mistral, DeepSeek are models: pattern-matching machines that predict the next piece of text. A model on its own has no interface and no account. It has to be hosted and served before anyone can use it.

A product is a model with everything around it that makes it usable. ChatGPT, Claude.ai, Gemini, Perplexity, Copilot, Cursor, Notion AI. Settings, chat history, file uploads, accounts, billing, connections into other software. Using ChatGPT means using a product that happens to run on a model.

A platform is direct access to the model through an API, sold to whoever is building their own. OpenAI’s and Anthropic’s APIs, AWS Bedrock, Azure OpenAI, Google Vertex AI. No interface and no chat history: a request goes in, an answer comes back, and everything a person eventually sees is built by whoever is calling it. The large cloud platforms carry open-weight models alongside closed ones.

Products sit on top because they are what people open. The models they wrap sit in the middle. Platforms sit underneath, serving those same models to anyone building. The names on each row are examples, not the full list.

Why the distinction matters

When people compare ChatGPT, Claude.ai and Gemini, they are comparing products. Each runs on a different model, but much of what makes one feel better than another is the product layer: chat history, file uploads, projects, workspaces, search, the mobile app. On everyday work the models are close enough that the wrapper decides the experience.

When engineers compare GPT against Claude, they are comparing models at the API level, with the product wrapper gone. Comparing a family’s top tier against its cheaper workhorse tier is still a model comparison, at a finer grain. The model families chapter has the tiers and what they cost.

When IT compares OpenAI direct against Azure OpenAI against AWS Bedrock, it is comparing platforms: different places to run the same or similar models, differing in how the data is handled, which compliance certifications they hold, who sends the bill and how they fit the cloud stack already in place.

The same model shows up at all 3 layers. GPT is a model family, ChatGPT is a product that runs on it, Azure OpenAI is a platform that serves it. “We use GPT” does not say which. Gemini is the sharper case: one word names both Google’s model family and the product people open.

The same job at all 3 layers

One job can usually be done at any of the 3 layers, and the layer decides what the purchase looks like. Licences are live the same afternoon. A build is weeks of someone’s time before anyone gets anything. Products are billed per seat, sometimes with a charge on volume on top. Platform access is billed per request, on top of a build that happens once. Leaving costs different amounts too: changing which model a build calls is a small change in the company’s own code, while moving off a product means leaving the history, the settings and the habits built up inside it.

Example

Let’s say a sales team wants help drafting emails. ChatGPT Team licences for everyone is the product layer, off the shelf. A sales tool inside the CRM that drafts the email where the rep already works is the product layer too, built for one industry and running on somebody else’s model underneath. A custom internal tool calling the OpenAI or Anthropic API is the platform layer, built in-house. That same tool hosted on Azure or Bedrock, so IT controls where the data goes, is the platform layer again, with the company’s own governance around it. All 4 are using AI, and finance is asked to compare them on one page.

When the conversation moves between the layers without naming them, decisions get made on mismatched assumptions: IT is thinking platform, sales is thinking product, and finance is comparing the two on price as though they were the same purchase.

Most AI products run on somebody else’s model

A large share of the AI tools landscape, including most of the “AI for [industry]” pitches arriving in inboxes, is built at the product layer on top of somebody else’s model. The vendor is not training a model of their own. They are calling the same APIs an in-house team would call, with their own wrapping around it.

That wrapping can be substantial: a workflow that fits a specific job, integrations into the systems already in place, prompts and checks written for the domain and tuned across many customers, a support relationship, an interface built for people who will never see an API. For many jobs the wrapping is most of the value.

It can also be thin: a generic chat interface, a prompt template and a markup on the underlying API price. Both exist in the same category and often look alike from the outside. What separates them is factual and checkable: whether the product does anything the platform alone doesn’t, what happens to the data on the way through, and whether the model underneath can be changed when a better one ships. What gets built around the model is the rest of that check: where the answers come from, what the product can act on, and what is verified before an answer goes out.

The layers are not a ranking. Most operations end up with a mix: products for jobs that already fit off the shelf, platform builds for the workflows specific enough to be worth the effort, and often an industry tool somewhere in the middle. Which layer a decision sits at is the first of the four questions every working AI system has already answered, and how the pieces fit together puts all four in one picture.