The main model families
Models come in families. Each family has several members, usually a larger, more capable one and smaller, quicker ones.
| Family | Made by | How you usually reach it |
|---|---|---|
| Claude | Anthropic | Hosted API |
| GPT | OpenAI | Hosted API |
| Gemini | Hosted API | |
| Grok | xAI | Hosted API |
| DeepSeek | DeepSeek | Hosted API; some models have open weights |
| Qwen | Alibaba | Hosted API; some models have open weights |
| GLM | Z.ai (Zhipu AI) | Hosted API; some models have open weights |
| Kimi | Moonshot AI | Hosted API; some models have open weights |
| Mistral | Mistral AI | Hosted API; some models have open weights |
| Llama | Meta | Open weights, run locally or through hosting providers |
Hosted means the model runs on the provider's servers and you pay for usage. Open weights means the model file is published, so you or any hosting company can run it. Which members of a family are open, and under what licence, is for the maker's own page to say. A router such as OpenRouter gives access to many families through one account.
Strong and slow, or fast and cheap
This is the most useful distinction. A large model reasons better across many files, plans multi-step work and recovers from its own mistakes more often. It also takes longer and costs more per request.
A small model answers quickly and costs little. It is good at contained jobs: rename this, write a test for that function, fix this error. Given a vague, sprawling task, it is more likely to drift.
Cost is counted in tokens, which are pieces of words. You pay for the tokens you send and the tokens the model writes. Agents send a lot, because they reread files as they work, so price differences add up quickly.
Context size
The context window is how much text a model can consider at once: your request, the files it has read, and the conversation so far. A larger window lets an agent keep more of a project in view. It does not guarantee the model uses all of it well, and filling it costs more. For big codebases, context size matters; for a small app, almost any current model has enough.
Tool use
Tool use, also called function calling, is the ability to ask the editor to do something: read a file, run a command, open a browser. An agent is built on it. A model that cannot call tools reliably can chat about code but cannot work on your project. This matters most with small local models, some of which do not support tools at all.
Vision
A model with vision can read images. For coding, that means you can paste a screenshot of a bug or a design mock-up, and an agent can look at a screenshot of the running app to check its own work. If you build interfaces, pick a model with vision for at least the design and testing steps.
Thinking levels
Many models can reason privately before they answer, often called thinking or reasoning. Tools usually expose this as a level. More thinking helps with planning, hard bugs and design decisions. It adds time and tokens, so it is wasted on a one-line change. Turn it up for hard problems and down for routine ones.
Why mixing models per task makes sense
No model wins on everything, and the leading one changes often. A practical setup matches the model to the job:
- Planning and architecture: a strong model with thinking turned up.
- Routine edits and tests: a fast, inexpensive model.
- Interface work: a model with vision.
- Private or offline work: a local open model.
The honest way to find out which model suits your project is to try the same task on two and look at the results. Comparing models side by side explains how to do that fairly. If cost is your main concern, see free and local models for coding.
Choosing models in VibeCode.surf
VibeCode.surf, a desktop AI code editor in beta, is built for mixing. You connect your own keys for Claude, OpenAI, OpenRouter, Gemini, Groq, DeepSeek, Z.ai (GLM), Mistral or xAI, or local models through Ollama and LM Studio, and give each agent a different model. The picker shows each model's price where the provider publishes it, every reply shows its tokens and cost, and you set a thinking level per agent from Off to Max. Compare mode runs one request on two models so you can keep the better result. The bring-your-own-key guide explains the setup.
It is an agent-first editor and does not yet offer tab completion or inline AI edits, so the choice of a model for as-you-type suggestions does not apply here.