Model usage is billed separately from agent slots. You have two options, and they are mutually exclusive per provider.
Prepaid credits
Buy credits up front; model calls draw them down. Coroid adds no markup to provider token rates, so a credit is worth the same as spending directly with the provider.
Top-ups come in fixed denominations. When credits run out, new model calls stop — your slots, projects, settings and history are unaffected, and adding credits resumes work immediately.
Bring your own key
Supply your own provider credentials and usage is billed by that provider, directly to you, at whatever rates you have negotiated. Coroid never sees the invoice.
BYOK is the better choice if you already have provider agreements, committed spend, or a procurement requirement that the model vendor be a direct supplier.
Keys are encrypted at rest, never returned by the API, and never included in settings exports — an export carries a placeholder and you re-enter the secret when importing elsewhere. See Provider keys.
What actually drives spend
Two things, in this order:
Which profile the work runs under. A profile carries a model for every agent, so moving work to a more capable profile raises the cost of every stage at once — and the developer stage runs most often, so it dominates the total.
How much context each task carries. Context is most of the token cost. Progressive skills keeps this bounded by loading only what the work needs — see Progressive skills.
Task count matters less than people expect. Ten small, well-specified tasks frequently cost less than one sprawling one, because the sprawling task explores more to work out what it should be doing.
Keeping it predictable
- Reserve the expensive profiles for work that warrants them, rather than defaulting everything to Smartest. See AI profiles.
- Write tight specifications with explicit out-of-scope lines. Exploration is what costs money. See Specifications.
- Watch per-project spend rather than one organization total — see Usage and reports.
Carbon
Coroid reports estimated operational emissions for inference, and the Low carbon profile prefers smaller models on lower-intensity grids. The estimate covers inference only — it is not a carbon-neutrality claim. See Carbon footprint.