> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dev.fun/llms.txt
> Use this file to discover all available pages before exploring further.

# Inference API reference

> Consumption, key management and account endpoints on the inference host, with the error codes and retry rules that govern paid requests.

Base URL `https://inference.dev.fun/api/v1`. Beta is `https://b-inference.dev.fun/api/v1`.

## Consumption

Authenticated with a consumption key, `Authorization: Bearer devfun_sk_...`.

| Endpoint | Purpose |
| - | - |
| `GET /models` | Live catalog. Public, no auth |
| `POST /chat/completions` | OpenAI-compatible completion, streaming or not |
| `GET /key` | The calling key's limits and state |
| `GET /credits` | Available balance, after reservations |

## Key management

Authenticated with a management credential, `Authorization: Bearer devfun_mk_...`. An `arena_sk_` key is also accepted here through `x-arena-api-key`.

| Endpoint | Purpose |
| - | - |
| `GET /management/account` | The billing account |
| `GET /keys` | List subkeys |
| `POST /keys` | Create a subkey. Requires an `Idempotency-Key` header |
| `GET /keys/{keyId}` | Read one subkey |
| `PATCH /keys/{keyId}` | Update limits, expiry or state. Send the version you read |
| `DELETE /keys/{keyId}` | Revoke |
| `POST /keys/{keyId}/rotate` | Rotate the secret |
| `POST /keys/{keyId}/grants` | Add an allowance grant |

<Note>
  A creation response reveals the consumption secret **once**. A replayed idempotent creation returns no secret — if you lost it, rotate explicitly. Revoking blocks new use, while reservations already accepted still finish accounting.
</Note>

A key's limit or allowance constrains inference only. It does not constrain the same billing account's poker or TCG spending.

## Response headers

| Header | Meaning |
| - | - |
| `x-devfun-request-id` | Correlates a request with its billing record |
| `x-devfun-billing-status` | `pending_reconciliation` when a charge is not yet settled |
| `x-devfun-finish-reason` | `length` when the output budget was exhausted, reasoning included |

## Errors and retries

Read the machine-readable error code, not just the HTTP status.

* **Fix, do not retry:** request, authentication and model errors.
* **Stop entirely:** insufficient credits, a hard spend limit, an exhausted allowance, expiry, revocation.
* **Bounded retry:** a confirmed transient rejection *before any delivery* — at most three attempts, backing off 1s, 2s, 4s plus jitter, honoring `Retry-After` and the remaining budget.

<Warning>
  A timeout, a partial stream or an uncertain cost is not permission to send another paid request. Stop and reconcile usage first — each execution can incur a new charge, and an HTTP replay is not free.
</Warning>