Dify OpenAI-API-compatible model setup

• A Dify workspace, cloud or self-hosted, with the OpenAI-API-compatible provider installed from Integrations > Model Provider.
• A RelaySwitch key. Sign in with an email code at the console and create one. Keys start with sk-relay-.
• A balance. Top-ups start at $10, by card or crypto, with no subscription. Before the first top-up the console chat runs 12 free turns a day for comparing models.

1. Open Integrations > Model Provider. If the OpenAI-API-compatible card is not there, install it under Install model providers.
2. Click Add Model on the card. Dify's docs describe this as giving the model “a name and its credentials”.
3. Fill the form:

• Model Name: the id exactly as the Models page prints it. Dify sends it as the model field.
• API endpoint URL: https://api.relayswitch.io/v1, with the /v1. The plugin README's rule: use the /v1 URL for LLM and chat models, and the bare origin for embedding, rerank, speech and transcription models, because for those types the plugin appends the API version itself. The RelaySwitch gateway also answers a doubled /v1/v1 path (checked 2026-10-31), so a wrong choice still works; the README's rule avoids it.
• Completion mode: Chat.
• Model context size and Upper bound for max tokens: the figures from the model's page. Dify uses them to cut history and cap output; set them lower than the model's if you want a tighter cap.
• Streaming: supported. The API streams server-sent events when a request sets stream to true, which Dify does for chat.

Save, then repeat Add Model for each id below. One key covers all of them; Dify stores it per model entry. Under Default Models at the top right, pick one of them as the system reasoning model so new apps start on it.

Prices per million tokens, context windows and the output cap, read from the live catalog when this page was built. The last column is one run of 2,000 tokens in and 500 out.

The Claude ids answer the OpenAI format here, so they go in through the same provider as the rest; no second provider is needed.

A chatflow with one LLM node on Claude Sonnet 5.5, run 500 times a day at 2,000 tokens in and 500 out each: $0.0039 a run, $1.95 a day. The same day on GLM 5.3 is $0.585.

A workflow with several LLM nodes bills each node as its own request, and a knowledge-retrieval step puts the retrieved chunks into the input of the node that reads them. Each request shows on the console's Usage page with its tokens and charge, so the split per node is readable there. See how pricing works for cached input and the rest of the billing rules.

From the machine that runs Dify, list the models the key can call:

Then send one streamed request the way an LLM node does. -i prints the headers, including the request id.

If both work and the Dify node does not, the problem is in the model entry. The API docs list every endpoint and error.

• Saving the model fails: Dify validates the entry with a request. A 401 means the key; a 404 means the Model Name is not a catalog id (a display name, or a typo).
• 402 in a run: the balance is too low. Top up in the console and the next run goes through.
• 429: the key hit its rate limit. Every app in the workspace shares the key you entered; give a busy app its own key with its own limits from the console.
• Output is cut short: Upper bound for max tokens is set below what the answer needs. Raise it up to the model's cap in the table.
• A /v1/v1 in an error message: the endpoint URL was entered with /v1 on a model type that appends it. The gateway still answers; for a clean path use the origin for those types, as the README says.

The menu path and the Add Model step are from Dify's Model Providers doc; the fields, the /v1 rule and the version (0.0.73 on the marketplace page) are from the OpenAI-API-compatible plugin README. Both read 2026-10-31. Prices are from the live catalog at build time.