The free tier runs on a managed model and is on by default. When you want a specific model or heavier usage, bring your own key: any OpenAI-compatible endpoint works, cloud or local.
The model must have strong native tool-calling (function-calling). Models without it cannot execute hardware actions or debug workflows.
Recommended models
| Model | Best for | Notes |
|---|
| Claude Haiku 4.5 | First try and production | The model we have IoT-specific benchmark evidence for. |
| DeepSeek-V4-Pro | Cost-sensitive long sessions | Larger context window means fewer overflow failures on long debug loops, and a lower price per million tokens. Via OpenRouter or the DeepSeek API. |
| GLM 5.1 | Worth watching | Previously our cost recommendation; DeepSeek-V4-Pro has since outpaced it on performance, context, and price. Still works as an OpenAI-compatible endpoint. |
| Setting | Value |
|---|
| API provider | OpenRouter (or any OpenAI-compatible endpoint) |
| Model | anthropic/claude-haiku-4.5 |
| Enable thinking | Off |
| Prompt caching (Advanced) | On |
