Think about hiring people. Two can both be smart and still differ: one is fast with routine admin, one is better at strategy, one is careful with legal language. Models are the same, all capable but not all good at the same work. Claude, GPT, Gemini, and other families differ in strengths, speeds, costs, and habits, and even inside one family a stronger model may be slower and pricier while a lighter one is faster and good enough for repeatable work.
How it shows up
A model is the capability layer, not the whole AI system. ChatGPT is an app, Claude Code is a harness, Codex is a tool for agentic coding. Inside those products, one or more models do the actual reasoning; the product decides how the model is wrapped, what tools it can use, and what rules it must follow.
This helps you choose. Ask a strong model to explore a messy new problem; once you’ve turned the work into a repeatable skill or process, a faster, cheaper model tier may be enough. The pattern we use: have the stronger worker figure out the work once, then write the instructions so the faster worker can repeat it.
You’ll also hear large language model, inference, and fine-tuning. An LLM is trained around language, inference is the model doing the work in the moment, and fine-tuning changes a model through training, which differs from giving it documents in context. When someone says “we trained it on our proposals,” they usually mean they gave it access to examples, which is useful but isn’t training.
Why you care
You rarely need a new model, just the right model, context, and instructions. The model is the worker’s capability, but the work still depends on the instructions, examples, and tools around it.