How it works · Model routing
Model routing.
Every message you send is automatically routed to the best available model for that type of task. You never need to think about this. Here's what's happening.
How routing works
Continio analyses each message before it reaches any model, scoring it across several dimensions: task complexity, emotional context, whether files are attached, whether the question needs current information, and more. This score determines which model handles the response.
This happens on every message, in milliseconds. The routing decision is logged — you can see which model handled a response by the label under the Continio logo in each message.
Which models are used
| Model | Used for |
|---|---|
| Claude Sonnet Anthropic |
Default for most things: coding, reflective questions, content creation, planning, emotional or personal conversations. Chosen because it respects continuity rules and tone better than alternatives. |
| GPT-4o OpenAI |
Vision tasks (any message with an image attached), complex analysis, and when GPT-4o is manually selected. |
| GPT-4o mini OpenAI |
Fast, lightweight responses: simple lookups, quick factual questions, background tasks like thread titles and suggestions. |
| Claude Haiku Anthropic |
Background extraction tasks: anchor detection, title generation, pattern analysis. Not used for conversational responses. |
| Grok xAI |
Manual selection only. Not in automatic routing. |
Routing modes
Routing assigns a mode alongside a model. The mode shapes how the model responds:
| Mode | What it means |
|---|---|
| reflective | Personal, exploratory, or emotional questions. Memory is weighted heavily. Tone is warmer. |
| practical | Task-focused, direct questions. Concise responses, less emotional framing. |
| coding | Code and technical tasks. Higher output token limit. File tools enabled. |
| content_creation | Writing, drafting, structured documents. File output enabled. |
| planning | Strategy, roadmaps, structured thinking. |
Manual override
You can override automatic routing by tapping the model selector next to the message input. The chosen model is used for that message and all subsequent messages in the session.
If you attach an image while using a model that has no vision capability (Grok, Haiku, or GPT-4o mini), Continio will automatically switch to GPT-4o for that message to handle the image. Your manual selection resumes after.
The automatic router is calibrated specifically for Continio's use cases. Manual override is useful when you have a strong preference for a particular model's style, not because automatic routing is unreliable.
Web search
Continio has live web search built in. It runs automatically when a message would benefit from current information — prices, news, links, recommendations, anything where training data might be stale. You don't need to ask for it.
If you ask for links (to buy something, find a place, look something up), Continio will search and return actual links rather than substituting general instructions.
Web search does not run when an image is attached — it would interfere with the model looking at the image.
Routing and the environment
Routing smaller tasks to lighter models (GPT-4o mini, Haiku) rather than always hitting the largest model is a deliberate design choice. It reduces per-message compute cost and makes Continio faster. It also reduces energy use per conversation.
See the Environment page for more on how Continio thinks about this.