Core Concepts
Definitions of groups, account pools, and inference credits, and the stages at which a request can be rejected.
Terminology
| Term | Meaning |
|---|---|
Inference credit 🍥 | The billing unit; account balance and subscription allowances are both measured in it |
| Group | A set of models with its own multiplier and capacity |
| Multiplier | The factor a group applies to a model's base price |
| Account pool | The set of upstream accounts a group maps to |
| Scope | Whether a key belongs to an individual or a team, which determines who pays |
| Composite key | A key bound to several groups, distinguished by prefix |
Groups
A group determines the available models, the billing price, the protocol format it accepts, and the upstream accounts that serve the request.
An API key must be assigned to a group at creation. A regular key binds to exactly one group; use a composite key to use several at once.
The multiplier applies to a model's base price. Prices shown in the model marketplace already include it and are the final billing prices.
A group is a property of an API key and is unrelated to the account's subscription plan (Lite, Plus, Pro, and so on).
The Claude Max group accepts only Claude Code clients. Other clients receive 403 this group only allows Claude Code clients. Check client restrictions in the model marketplace before choosing a group; see The Claude Max Group.
Account Pools
An account pool is the set of upstream accounts a group maps to. The Pro pool includes Pro models and may receive higher priority when the service is under load; otherwise it is no different from the Plus pool.
Account type determines whether certain features work. Forcing Fast Mode on requires an OpenAI or Anthropic platform account, and Anthropic accounts must use API key credentials; Bedrock, Vertex, and OAuth credentials are skipped silently.
API Keys
Scope determines who pays, cannot be changed after creation, and is switched from the top right of the API keys page.
| Scope | Paid by | Available groups come from |
|---|---|---|
| Personal | You | Your own group entitlements |
| Team | The team owner | The owner's group entitlements |
A regular key binds to one group at creation, and model IDs are written as-is.
Enabling Composite Key at creation binds 1 to 20 groups, each with its own prefix, and the group is selected per request via prefix/model-id. A single key handles requests across multiple groups.
Cost Calculation
A request is billed as input, output, and cache, each at the rate of the group the request actually hit. For composite keys, that is the group matched by the prefix.
Subscriptions are consumed before balance: the allowance expiring soonest is used first, and the account balance is charged only when no subscription is available.
Fast Mode uses the high-priority tier and costs roughly twice the regular rate.
Where a Request Is Rejected
When a request fails, check in the following order:
| Stage | What to check |
|---|---|
| Key | Exists, not disabled, not expired. Team keys also require both the member and owner accounts to be active and the team not suspended |
| Group resolution | Composite keys need a prefix on the model ID, and the prefix must exist in the mapping, otherwise COMPOSITE_KEY_* is returned |
| Model permission | Whether the model belongs to that group |
| Endpoint support | Composite keys do not support WebSocket / Realtime endpoints |
| Quota | Account balance, subscription allowance, and a team member's daily / weekly / monthly limit - any shortfall rejects the request |
| Upstream | Only after all of the above does the request reach the model provider |
TokenFlux does not offer embedding models.
For the status code and message behind each case, see Error Codes.
Related Pages
- Troubleshooting - locate a problem by symptom when a request fails
- Create API Key - pick a group and generate a key
- Billing - multipliers, subscriptions, and top-ups
- Composite Key - use several groups from one key
- API Endpoints - the address your client needs