Choosing models
A model is the SI brain an agent thinks with. C5 does not lock you into one. You can use several, and let different jobs use different ones.
Why this matters #
Models differ in three ways that pull against each other:
- How capable they are. Stronger models handle harder problems.
- How fast they are. Smaller models answer quicker.
- What they cost. Stronger usually costs more per use.
Using the strongest model for everything works, but you pay for power you do not need on simple jobs. Using the cheapest for everything saves money until it gets something important wrong.
The answer is to match the model to the job.
A simple approach #
| Kind of work | What to use |
|---|---|
| Sorting, labeling, simple formatting | A small, fast model |
| Everyday writing and summarizing | A mid-range model |
| Hard reasoning, tricky analysis, planning | Your strongest model |
| Bulk repetitive work | A local model, if you can |
C5 sets sensible defaults so this works reasonably well before you touch anything.
Changing the defaults #
- Go to Settings.
- Open the model routing section.
- Choose which model handles which kind of work.
- Save.
You can also set a single model for everything if you prefer simplicity over savings.
Running models on your own machine #
You can run models locally instead of paying a service. Local models:
- Cost nothing per use once set up
- Keep everything on your computer — nothing leaves at all
- Keep working offline
- Are usually less capable than the best paid models
- Need a reasonably powerful computer
A common setup is local models for routine work and a paid model for the hard parts. See Model providers for how to connect one.
When a model is unavailable #
Providers have outages. When one is down, C5 can fall back to another model rather than stopping. You choose the order it tries, in Settings › Integrations › Coordinator Priority (with its own column for chats, tasks and C5's background work).
That order holds among the coordinators that are working. One that keeps failing is tried last, until it recovers:
- when 3 of its last 5 answers came back as an error or empty, or
- when 3 of its last 5 attempts got a server error (5xx), a refused connection, or a gateway that stayed dead.
Being tried last lasts a minute at first, then twice as long each time it fails again once let back in, up to 15 minutes. A good answer resets that, and so does restarting C5. It is never dropped from the list: if everything else fails, it is still tried. Its card in Settings › Integrations says which: Skipped — repeated failed generations for empty or broken answers, or Tried last — repeated gateway errors for server errors, refused connections and a dead gateway (see When a provider fails).
Cancelling a task is not a failure: C5 stops there and tries no other model.
The order matters for one more thing: how many task steps may run at once. Cloud Concurrent Agents in Settings › General only adds running agents while a paid provider is first in the tasks column; see Settings.
This is worth setting up if you rely on schedules that run while you are not watching.
Watching the cost #
The Ops charts break spending down by model. If one line dominates, that is where to look first — often a job is using a stronger model than it needs.
Tip Change one thing at a time and check Analytics afterwards. If quality holds steady on a cheaper model, keep the saving. If it drops, put it back.
