# Credits & usage

> Credits are the unit of AI usage. Each reply costs credits based on its model; your plan includes a monthly allotment. Learn what spends credits, how per-model costs work, and how to track and stretch your balance.

A **credit** is the unit of AI usage in Bookbag. Every AI reply your agents generate consumes credits, and each plan includes a monthly credit allotment. This page explains exactly what spends credits, how per-model pricing works, and how to keep usage under control.

> **CREDITS APPLY TO THE HOSTED CLOUD SERVICE:** Credit metering applies to Bookbag Cloud. The self-hosted open-source build is unmetered — usage is unlimited and credits aren't enforced.

## What costs credits

Credits are spent on AI replies, and only on AI replies. Building and maintaining your agent is free.

| Action | Costs credits? |
| --- | --- |
| An agent generating an AI reply | Yes — by the reply's model credit cost |
| Adding or retraining a data source | No |
| Crawling, chunking, and embedding | No |
| Inspecting the prompt in the playground | No (it doesn't call the model) |
| Replies on the Bookbag Local model | No (1-credit cost, and not provider-backed) |

> **TIP:** Because training is free, you never have to ration how much knowledge you give an agent. Add sources liberally and retrain whenever your content changes.

## Per-model credit cost

A reply's cost equals the **credit cost of the model** that produced it. Cheaper, faster models cost less per reply; premium models cost more. This is the single biggest lever on your usage.

| Model | Credits per reply |
| --- | --- |
| Bookbag Local | 1 |
| GPT-4o mini | 1 |
| Claude Haiku 4.5 | 2 |
| Claude Sonnet 4.6 | 4 |
| GPT-4o | 5 |

A workspace running everything on GPT-4o spends up to 5x the credits of one running on GPT-4o mini for the same conversation volume. See [Models & model choice](/docs/agents/models) for how to pick the right model for each agent.

## Plan allotments

Each plan grants a monthly credit allotment when you subscribe. New workspaces are provisioned with their plan's credits immediately so you can build and test right away.

| Plan | Credits / month |
| --- | --- |
| Free | 100 |
| Hobby | 2,000 |
| Standard | 12,000 |
| Pro | 40,000 |
| Enterprise | Custom |

See [Plans & billing](/docs/workspace/billing) for full plan details and pricing, or the [pricing page](/pricing).

> **RUNNING OUT OF CREDITS:** In cloud mode, when your balance reaches zero, AI replies are blocked until your allotment renews or you upgrade. Watch your balance as you scale — high volume on a premium model can draw down an allotment fast.

## Tracking usage

Bookbag logs every metered reply to a usage ledger, so you can see where credits go:

- **Per-agent credit usage** appears in [Analytics](/docs/agents/analytics), windowed by date — so you can see which agent and which period drove spend.
- **Workspace balance and total usage** are summarized in [Plans & billing](/docs/workspace/billing).
- Each ledger entry records the model, agent, and conversation, so usage is fully attributable.

## Stretching your credits

1. **Use the cheapest model that's accurate** — For most ecommerce FAQ the data does the work — a 1-credit model answers as well as a 5-credit one. Reserve premium models for agents that truly need them.
2. **Pin high-frequency answers as Q&A** — Q&A pairs still cost a reply, but precise pinned answers reduce re-asks and escalations, lowering total volume.
3. **Let the agent resolve more** — A higher resolution rate means fewer back-and-forth turns and fewer escalations — each saved turn is a saved reply.
4. **Consider BYO at scale** — High-volume teams with their own provider accounts can [bring their own keys](/docs/agents/byo-keys) to move LLM cost off credits and onto their provider bill.

## Common questions

**Do credits roll over month to month?**

Allotments are granted per month against your subscription balance. Treat your monthly allotment as your working budget and size your plan to your expected volume.

**Does a multi-turn conversation cost one credit or several?**

Credits are charged per AI reply, by the model's cost. A conversation with several agent replies spends credits for each reply, so resolving in fewer turns saves credits.

**How do I avoid spending credits while testing?**

Use Inspect prompt in the [playground](/docs/agents/playground) to see the assembled prompt without calling the model — it's free. Live test chats do generate replies and cost credits.

**What if I want unmetered usage?**

Two options: self-host the open-source build (unmetered), or [bring your own keys](/docs/agents/byo-keys) so provider-backed replies bill to your provider instead of your credit balance.

## What's next

- [Models & model choice](/docs/agents/models) — Per-model costs and how to pick the cheapest accurate model.
- [Bring your own API keys](/docs/agents/byo-keys) — Move LLM cost off credits onto your own provider account.
- [Analytics](/docs/agents/analytics) — See per-agent credit usage over time.
- [Plans & billing](/docs/workspace/billing) — Plan allotments, pricing, and your balance.