The model picker
Under the chat box, next to the mode. Each chat keeps its own model. The menu groups models by maker, recommended ones first, with Show more for the rest, and each line says what the model is for and what it uses. Settings, AI, Default Model sets the model new chats start with (Claude Sonnet 5 out of the box).
On Starter, the models included on Piye credits are listed under On your Starter plan; the others say what opens them (Pro or API key). See AI models.
Effort
Effort sits beside the model, for models that think before they answer. It sets how long the model reasons:
| Level | When to use it |
|---|---|
| Low | Quick answers and small, clear edits |
| Medium | Everyday work |
| High | Hard problems, careful changes |
| Max | The hardest work, where depth matters more than speed |
Higher effort thinks longer and uses more credits. Each model offers only the levels it supports, and starts at its own standard level. With Claude's latest models, changing effort mid-chat doesn't make the conversation more expensive to continue. The level used shows on each reply.
Where the AI runs
| Route | Who pays | What the agent can do |
|---|---|---|
| Piye credits (signed in) | Your plan's credits | Piye's full agent: every tool |
| Your own API key | Your provider | The standard IDE tools: files, terminal, editor and chat |
| Local models (Ollama) | Nothing | The standard IDE tools |
When you're signed in, credits come first and a key covers the models your plan doesn't include. Only API keys (Settings, AI) runs every Code request on your keys instead. Piye never moves work to your key without asking. See Your own API keys and local models.
What a reply costs
Under each reply: the credits it used (on Piye credits) or what it cost on your key ("$0.42 on your key"). The account menu shows the last 14 days of use, and the web console's Usage page shows every call. See Credits and usage.
To spend less: start a new chat for a new task (long chats are re-read on every reply), pick a smaller model for small jobs, and keep effort at the model's standard level unless the task needs more.