Codex, Claude, Llama and Ollama explained for beginners
Understand the difference between an AI model, an assistant, a local model runner, and an API before deciding what to use or pay for.
AI names can make shopping for help confusing. You hear about Claude, Codex, Llama, and Ollama as though they were four versions of the same thing. They describe different parts of a system.
Before comparing costs, ask three questions: what generates the answer, what software helps it do the work, and where does that work run?
A model generates the answer
A model is the trained system that interprets input and produces an answer. Different models have different capabilities, sizes, and operating requirements.
Llama is Meta's family of AI models. It is not the name of a single chat application. Particular Llama releases have their own model sizes, capabilities, and licenses. A model that someone can download is not automatically small enough for their laptop. See Meta's Llama model repository (opens in a new tab).
Claude is Anthropic's AI model family and the name used for its assistant products. That shared name can cause confusion: choosing a Claude model and choosing the application through which you use Claude are related decisions, but they are not identical.
An assistant connects a model to a workflow
An assistant application gives you a place to ask questions. An agent can also use permitted tools to perform tasks, such as reading project files or running checks.
Codex is OpenAI's coding agent. You can ask it to work on software tasks, investigate a problem, or review changes. Its model and access route affect what it can do and how usage is accounted for. A Codex task can involve multiple steps rather than a single answer. See OpenAI's documentation (opens in a new tab).
Claude Code is Anthropic's coding assistant. It can inspect and change project files and use development tools, subject to its permissions and environment. It is different from simply pasting a question into a general chat. See Claude Code's overview (opens in a new tab).
Neither product name establishes which assistant will finish your particular task most cheaply. Check the available model, billing route, and actual results.
A model runner makes models available
Ollama is software for running and using supported models. It can run models on your own hardware, and it also offers cloud capabilities. Ollama and Llama are different names: one is a runner and service ecosystem; the other is a model family. Ollama supports models beyond Llama. See Ollama's documentation (opens in a new tab).
llama.cpp is another name you may encounter: software used to run compatible models. It is not a Llama model release. You do not need to understand every runtime to begin using an assistant. See the official llama.cpp project (opens in a new tab).
Local and cloud describe where processing happens
Local inference means your computer performs the model's answer-generation work. It uses your hardware, memory, power, and time.
Cloud inference means a provider's computers perform that work. Using an application installed on your computer does not, by itself, make inference local.
An Ollama workflow can use local or cloud models. Its official FAQ explains a local-only option. Check the selected model and destination rather than relying on an app icon. Ollama FAQ (opens in a new tab)
A larger workflow can mix locations. A local document search might send retrieved passages to a hosted answering model. Check the complete route when deciding where information may go.
Subscription and API describe access and billing
A subscription provides access under a plan's terms and usage limits. It should not be read as unlimited work or permission to connect any third-party tool.
An API is a software interface that lets an application request model work. API access may be billed separately from an assistant subscription. A connector accepting an API key does not mean your existing subscription pays its bill.
Tokens are small pieces of information a model processes. They are not a fixed number of words. Prompts, relevant files, conversation context, tool results, and generated answers can contribute to usage.
Read the names as a simple chain
| Part | Example | Question to ask |
|---|---|---|
| Model | A particular Claude or Llama model | Can it handle this task accurately? |
| Assistant | Codex or Claude Code | What tools and workflow does it provide? |
| Runner | Ollama | Which model is running, and where? |
| Access route | Subscription or API | Which allowance or bill does this use? |
You do not need one of everything. Begin with a task, the tools you already have, and a way to check success. Then choose the least expensive route that can reliably complete it.