AI credits
How Narratomi meters AI usage, which limits can stop a run, and what you see when they do.
AI work is metered in credits, not tokens. One credit is one tenth of a US cent of model cost, so a call on a small model costs a fraction of what the same tokens cost on a large one. Every model call is priced from its own input and output token counts and rounded up, with a floor of one credit per call.
Per-plan pools and caps are in plan limits.
What spends credits
| Event | Metered |
|---|---|
| A co-writer action (any of the five) | Yes, once per call |
| A co-writer repair retry after malformed output | Yes, both attempts count |
| A co-writer run you cancel mid-stream | Yes, for whatever was spent before the abort |
| A co-writer run that comes back empty | Yes |
| An agent chat reply | Yes, one call |
| An agent run: the planning call, each phase, each re-plan | Yes, every call |
| Reading a stored conversation, accepting, reverting, merging, discarding | No |
An agent run spends the credits of whoever started it, not of whoever reviews or merges it.
The two limits
Monthly pool. Your plan's credits for the calendar month. Purchased top-up packs add to it, and the meter counts the extended total.
Daily burst cap. A per-day ceiling that resets at midnight UTC. It is an abuse brake, so top-up packs do not raise it.
Both are checked before a model call, never during one. Hitting either produces a refusal rather than a partial answer.
Where the numbers are
- The co-writer bar shows
AI budget: N% usedagainst your monthly pool. It changes color once you pass 80%. - The agent thread's footer shows the run's spend against the starter's monthly pool.
- While a run is live, the bar shows
◆ agent run activeand the run's credits instead of the percentage. - Settings shows your plan's monthly credits and any purchased top-up balance, and sells packs of 1,000 credits.
What you see when you run out
The co-writer refuses before the call and the draft column explains why:
- Monthly pool exhausted: "You've used this month's AI credits (7500/7500). They reset October 1, or buy a credit pack." The message links to Settings. The date is the first of the next month (UTC), because the pool is counted per calendar month, not per billing period.
- Daily brake hit while the month still has room: "You've used today's AI credits, they reset September 12. This month you're at 1200/7500." The message links to the pricing page. Here the date is tomorrow, since the daily brake is what stopped you.
The agent shows the same wording as a toast and does not start. Asking it for a run while one is already going is a different refusal: "A run is already writing on this project. Wait for it to finish, or stop it, then send this again."
An agent run that runs out mid-run stops at its next phase boundary, exactly as a Stop would. Everything written up to that point is on the review branch and can be reviewed, rewound, merged, or discarded normally.
Model class
Some plans are limited to small models. When a co-writer action would need a standard model, it is refused with "This action needs a standard model; your plan runs mini models only." Generate branch runs on the larger model slot, so it is the action this usually affects.