Every month, most of us now pay for AI the way we pay for electricity: a standing charge that turns up on the card and gets waved through as the cost of being in business. But the bill isn't fixed. It's a setting you haven't touched, sitting behind a default someone else chose for you. And the only thing standing between you and a cheaper, better-fitting setup isn't technical skill, it's a handful of words the industry uses as if they were interchangeable and never bothers to explain.

So what follows is a way of seeing. There's a glossary at the bottom if you want one, but the value is in the picture itself. Once you've got it, you can't unsee it, and plenty of decisions that felt like guesswork start to feel like choices.

The setting you haven't touched

Start with the one you already have in front of you. Open Claude Cowork, or ChatGPT, or almost any AI product, and somewhere there's a small selector offering you a choice of model — Opus or Sonnet, say. Most people leave it on whatever it arrived set to, because the names mean nothing and the whole thing looks like something for the technical folk to worry about. It isn't. Opus is the powerful, slower, more expensive option; Sonnet is the lighter, faster, cheaper one. Same product, same buttons, same screen, same way of working; different model doing the actual thinking. For a hard piece of reasoning, reach for the powerful one. For drafting an email, summarising a document, the everyday back-and-forth, drop to the lighter one and your subscription stretches further. Nothing else changes. You've simply stopped paying for power you weren't using.

That small toggle is the whole idea in miniature. The model — the part that does the reasoning — is a separate thing from the product wrapped around it, and you can change one without changing the other. I made a version of this point eighteen months ago from the other side, arguing that most of what you're paying for lives in the product, not the model. This is the version you can act on without writing a line of code.

What you're actually paying for

Two things are doing the work whenever you open up an AI tool. There's the product — Claude Cowork, the ChatGPT app, the Codex tool, Cursor — and there's the model running inside it: Opus 4.8, GPT-5.5, Gemini 3.1 Pro. The bill bundles them. The marketing blurs them. But they're separate decisions made by separate parts of the same machine, and when the product is the kind that can run tasks and edit files rather than just answer questions, the trade calls that an agentic harness — a useful name because the harness is the layer that decides whether you can swap the model underneath.

None of the big sellers name these layers consistently. The cleanest example is Google: Gemini is the company's AI brand, the product you open, and the model line, so when someone tells you they "use Gemini" you've learned almost nothing about what they're actually running. OpenAI is barely tidier — "ChatGPT" is the product most people mean, and Codex is now both a product and a family of models. Anthropic happens to keep them separate, but that's a stylistic choice on their part, not a market norm. The cost of letting the marketing pick your vocabulary is that you can't tell which layer you're being sold.

When the bundle comes apart

Where the distinction earns its keep is the moment your tidy bundle starts to come apart, and how far it comes apart depends on what you've bought. Claude Cowork lets you change the model within Anthropic's own range, Opus for Sonnet. A tool like Cursor lets you pick across companies. And something like OpenClaw, or Claude Code itself, will run almost anything you point it at, including open-weight models such as Moonshot's Kimi K2.6 that you can host on your own machine. That last part is the proof the separation is real: people genuinely run Claude Code on a non-Anthropic model, and it still works. If the two were one welded unit, you couldn't. The harness is the thing you work in: the loop, the tools, the memory, the permissions. The model is the reasoning you plug into it. The bundle just decides how far you're allowed to reach for a different one.

There's even a whole class of tool whose entire reason for existing is to hand you that reach — OpenClaw, community-built and now moving to its own independent foundation, is the obvious one. It's a harness that will run whatever model you give it: bring your own. With Anthropic, the company building the product also builds the model and would, understandably, prefer you used both. With a bring-your-own-model tool, the product stays the product and you pick the reasoning yourself, or you let it be no one's, in the case of open models that belong to nobody. Whether that flexibility is worth the extra setup will depend on you; for most small teams, the truthful answer is they don't switch often enough to feel locked in either way. For a business thinking past this quarter, though, that's the un-locked-in path: you're not betting your whole workflow on a single vendor's roadmap and a single vendor's pricing.

Everything I've described is shifting under your feet. Codex used to be a model and is now also a product. OpenClaw was Clawdbot, then Moltbot, then OpenClaw, all in the space of a few months. The vocabulary keeps sliding because the industry hasn't settled on which layer it's actually selling, so the names chase wherever the money is rather than describing what's underneath. You're being asked to make purchasing decisions in a fog the sellers are busy generating.

The integration question

There's one more layer worth knowing about, and it's a buying question rather than a what's-inside-the-machine question. Can your AI actually touch your calendar, pull a document, post into Slack, edit a row in your accounting software? Or can it only describe those things from outside? The current standard for letting it do the touching is called MCP. You can cheerfully forget what the letters stand for; the only question that matters is whether your AI can connect to the other tools you already pay for. Without it, an assistant is clever but sealed off, able to talk about your work and never to touch it. With it, it can act on the real thing. So "do you support MCP?" is worth asking out loud when you're weighing up a new bit of kit, because the answer tells you whether your AI will be able to work with it or whether you've just bought an island. Whether MCP itself wins as the standard or another contender takes its place isn't yet settled; what's settled is that the connect-to-the-rest-of-your-stack question is the right one to be asking. I've written before about how whole categories of software are now being rebuilt around AI agents as their primary customer rather than humans; whatever the protocol ends up being called, it's the connector that decides whether the agent you've paid for can join in.

None of this asks you to write a line of code or read a single specification. It asks for one shift in how you look at the thing you're paying for: stop treating "an AI" as a single, indivisible object, and start seeing the model, the product around it, and its reach into your other tools as three separate choices that are yours to make. Do that, and the intimidating little toggle becomes a lever for your own budget, and the glossy product page becomes a short list of questions you now know how to ask. Learning to name the parts is how you stay the one making the choices.

A short glossary

Company: the organisation that builds and sells the technology. Anthropic, OpenAI, Google, Moonshot.

Brand: the consumer-facing name the company markets under. Claude, ChatGPT, Gemini. Often, confusingly, doubling as something else on this list.

Product: the actual thing you open and use. Claude Cowork, the ChatGPT app, Cursor, the Codex tool.

Agentic harness: a product that lets the AI take real actions (run tasks, use tools, edit files) rather than only answering questions.

Model: the part that does the reasoning. Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Kimi K2.6. Separate from the product wrapped around it, and often swappable.

MCP: the standard that lets an AI connect to your other tools and data. The thing that decides whether your assistant can act on your world or only describe it.