Get started

# Dagon Inference documentation

Dagon Inference is a metered gateway in front of upstream model providers. You call it with any OpenAI or Anthropic SDK, each request goes to the cheapest healthy provider serving its model id, and you pay that provider's token price plus 10% from a prepaid balance or per request in USDC. Every request gets an Ed25519 signed receipt.

**On this page**

- [Base URLs](https://inference.dagon.so/docs#base-urls)
- [How a request is handled](https://inference.dagon.so/docs#how-a-request-is-handled)
- [Where to go next](https://inference.dagon.so/docs#where-to-go-next)
- [Machine-readable files](https://inference.dagon.so/docs#machine-readable-files)

## Base URLs

| Client | Base URL | Endpoints |
| --- | --- | --- |
| OpenAI SDKs | `https://inference.dagon.so/v1` | `/v1/chat/completions`, `/v1/responses` |
| Anthropic SDKs | `https://inference.dagon.so` | `/v1/messages`, `/v1/messages/count_tokens` |

Authenticate with `Authorization: Bearer <key>` (Anthropic SDKs may send `x-api-key` instead). Keys start with `gk_` (prepaid). Without a key, the inference endpoints answer `402` with [x402](https://inference.dagon.so/docs/x402) payment requirements.

## How a request is handled

1. **Authenticate** the key, or price a keyless request for x402.
2. **Hold** the request's maximum cost against the balance: an upper bound of the prompt plus the full output cap, at the rates of the dearest route that may serve it, plus the markup. Short of funds: `402 insufficient_credits`, and nothing is sent upstream.
3. **Route** to the cheapest healthy provider serving the model id and endpoint, falling back to the next before any output.
4. **Capture** the provider's reported usage, release the rest of the hold, and sign a receipt naming the provider and price version.

Details: [Models and pricing](https://inference.dagon.so/docs/models), [Receipts](https://inference.dagon.so/docs/receipts).

## Where to go next

[**Quickstart** Create a key, fund it and make a first request.](https://inference.dagon.so/docs/quickstart) [**Chat Completions** The OpenAI endpoint most tools use.](https://inference.dagon.so/docs/chat-completions) [**x402 pay per request** Agents pay in USDC with the request, no key.](https://inference.dagon.so/docs/x402) [**Receipts** Check every charge against the published prices.](https://inference.dagon.so/docs/receipts) [**Errors** Every error shape and code.](https://inference.dagon.so/docs/errors) [**Agent setup** llms.txt, OpenAPI and the MCP server.](https://inference.dagon.so/docs/agents)

## Machine-readable files

- [`GET /v1/models`](https://inference.dagon.so/v1/models): the live, canonical model and price list. No key needed.
- [/models.json](https://inference.dagon.so/models.json) and [/models.md](https://inference.dagon.so/models.md): a build snapshot of the same list.
- [/changelog.xml](https://inference.dagon.so/changelog.xml) (Atom) and [/changelog.json](https://inference.dagon.so/changelog.json): every change to that list, dated: new, removed and renamed models, refusal labels, providers and prices ([read it here](https://inference.dagon.so/changelog)).
- [/openapi.json](https://inference.dagon.so/openapi.json): OpenAPI 3.1 description of the public API.
- [/llms.txt](https://inference.dagon.so/llms.txt) and [/llms-full.txt](https://inference.dagon.so/llms-full.txt): the API summarised for language models.

Not affiliated

Dagon Inference is not affiliated with OpenAI, Anthropic or the upstream providers. Compatibility refers only to request formats. It is operated by Inkwell Finance, Inc. d/b/a Dagon.

Last updated 2026-11-08

[Next **Quickstart**](https://inference.dagon.so/docs/quickstart)