LLM Provider vs Chat Tab: Who Sells Tokens, Uptime, and the Data Policy

“We use ChatGPT” usually mixes three things: a chat app a person opens, a model name (GPT, Claude, Gemini), and a company that meters tokens and sends an invoice. That company is the LLM provider. If you skip the distinction, you pick a tab instead of a vendor - and you cannot tell legal, finance, or a developer what you actually bought.

I already wrote what an LLM is and how chat plans versus API pricing compare. This piece is narrower: who stands between you and the model, what you buy from them, and why swapping the logo later hurts.

1. Four names people collapse into one

Keep them separate on a contract and in a repo. Mixing them is how a “ChatGPT integration” becomes an unmovable vendor.

• Model: the weights. GPT-4.1, Claude, Gemini, Llama, Mistral. A name you put in config. Not a company.
• Chat app: ChatGPT, Claude.ai, Gemini in the browser. A person types. Fine for drafts. Not how a website, a CRM, or a bot calls a model.
• API: the HTTPS door your product uses. Same model family, different contract, logs, and rate limits.
• Provider: who runs that door and bills you. OpenAI, Anthropic, Google, Mistral, or a host that serves someone else’s open weights (Groq, Together, Fireworks, Azure, AWS Bedrock).

2. What you actually buy from a provider

You do not buy “AI”. You buy a metered service with a policy. The invoice is the easy part.

• Tokens in and out, with a ceiling. Burst traffic from a campaign is a product incident, not a surprise invoice - if you cap it.
• A data policy: whether prompts are used to train, where they sit, how long they live. This is the question legal actually asks. The model name is not an answer.
• Uptime and a region. A bot that bills customers cannot shrug at an EU outage the way a person shrugs at a slow chat tab.

3. First-party vs a host of open weights

Two provider shapes. Confusing them is how you “save money” and lose the model you tested.

• First-party: OpenAI, Anthropic, Google, Mistral selling their own models. One throat to choke. The model and the bill share a brand.
• Host: Groq, Together, Fireworks, Azure OpenAI, Bedrock. They run a model (sometimes OpenAI’s, sometimes Llama). Faster or cheaper or in your cloud - and a second company in the chain when something breaks.

4. Switching models is cheap. Switching providers is not

A model id in config is a one-line change after you re-test quality. A provider is keys, billing, DPA, logging, and every prompt that assumed their tool-calling quirks. I keep a thin adapter: one interface, one place for keys, no OpenAI types leaking into the rest of the app. That is the work. Picking a logo on a slide is not.

Conclusion: name the vendor, not the tab

An LLM provider is who sells you tokens, a region, and a data policy. The chat tab is for people. The model name is a line in config. If you want a production bot, a CRM step, or a page that calls a model, you pick a provider and an adapter - not “we use ChatGPT.” I wire that into Next.js apps. Write via the form with whether this is a coworker tool or a product feature; we will pick a provider that can actually sit behind a webhook.