# How to choose a model

Which model to take for the task — a fast and cheap one or the smartest — and how not to overpay.

## Three questions before choosing

1. **How hard is the task?** Summarizing a note or making self-check questions — a fast model will
   cope. Program architecture, hard math, a long analysis — you need a strong one.
2. **How much text will the model read?** A big repository or a long document needs a large context
   window and costs more on every request.
3. **How much are you willing to spend?** Model prices differ by dozens of times. Start with a cheap
   one and move to a strong one only when the cheap one does not cope.

## Classes of models

| Class | What for | Price |
|---|---|---|
| Fast and small | summaries, questions, simple edits, drafts | the cheapest |
| General-purpose | most tasks: code, explanations, agent work in a repository | medium |
| Flagship and «thinking» | hard code, math, long reasoning | the most expensive |

Every major provider has models of all three classes — look for the names in the provider articles,
and compare prices in the [model catalog](https://skillok.org/docs/en/ai-models/models).

## How not to overpay

* **Cheap first.** A fast model is enough for most questions about your notes.
* **Short rules.** The `AGENTS.md` file goes to the model with every request — keep it short.
* **A repeated beginning.** Providers cache an identical beginning of a request (rules, tools) and charge
  many times less for it.
* **A limit on the connection.** The connection settings let you set a monthly limit in dollars — the
  model is not called beyond it.

## If a model is not available from Russia

OpenAI, Anthropic and Google do not sell access to their models in Russia directly. In SKILLOK their
models work through OpenRouter — one key for models of many companies. In detail —
[Connecting in SKILLOK](https://skillok.org/docs/en/ai-models/connecting-in-skillok).

---

https://skillok.org/docs/en/ai-models/choosing-a-model