Dify with Claude and GPT via an OpenAI-Compatible API

Dify is an open-source platform for building LLM apps: chatbots, agents, workflows and knowledge-base (RAG) apps. You can use Dify Cloud or self-host it with Docker.

Dify reaches models through model provider plugins installed from the Dify Marketplace. Each provider stores a key and an address for the whole workspace, and only workspace owners and admins can change them. The providers live on the Model Provider page: in current Dify Cloud it is under Integrations, in older self-hosted versions under Settings in the account menu. The field names below were checked in October 2026 against the Dify docs and the source of the official openai_api_compatible plugin (version 0.0.68) and anthropic plugin.

Start with OpenAI-API-compatible. You type each model ID yourself, so new models work as soon as they appear in the APIVAI list. The general details of both formats are on the OpenAI-compatible API and Claude API proxy pages.

1. Sign up with your email
2. Top up your balance, from $10, by card, crypto, Alipay or WeChat Pay
3. Create a key on the Dashboard and copy it. Give it a budget limit if several people build apps in the same workspace.

1. Open the Model Provider page.
2. In the list of providers you can install (or on the Dify Marketplace), find OpenAI-API-compatible and click Install.

On the OpenAI-API-compatible card, click Add Model and fill in the form:

Leave the other fields at their defaults. Older versions of the plugin label the address field API endpoint URL; the value is the same.

A few notes on these fields:

• Model Name is what Dify sends to the API, so it must be the exact ID. The display name is only for the interface.
• Model context size is the number Dify uses when it fits conversation history and retrieved knowledge into the prompt. A value below the model's real window is fine and keeps long chats cheaper.
• Upper bound for max tokens only sets how high the Max Tokens parameter can go in your apps. Thinking is on by default and its tokens count as output, so leave room above 4096.
• Function Call Type: Tool Call lets Agent apps use native tool calling, which APIVAI supports.

Click Save. Dify checks the credentials with a test request before it saves the model. Repeat Add Model for every model you want, for example gpt-5.5 and claude-haiku-4-5, with the same key and address.

• In an app, open the model selector, choose your model under OpenAI-API-compatible, and set Max Tokens to 4096 or more in its parameters.
• In a workflow, each LLM node has its own model, so cheap and strong models can sit in the same flow.
• Under Default Models on the Model Provider page, set the System Reasoning Model, the workspace's default model for general LLM tasks.

1. Install the Anthropic provider plugin and open its setup.
2. Set API Key to your APIVAI key and API URL to:

Do not add /v1: the plugin passes this address to the Anthropic client, which appends /v1/messages. Leave Enable request metadata on Disabled.

1. The plugin comes with a list of Claude models. Some IDs in it match APIVAI's, such as claude-sonnet-4-6; others use different or dated names. If a model is missing or returns "model not found", use Add Model on the Anthropic card and enter the exact ID from the APIVAI list, such as claude-opus-5-5.

Prices per 1M input / output tokens as of 2026-10-07:

The model list changes over time. Take IDs from the pricing page or from the API:

Two examples, with thinking tokens counted as output:

• Support chatbot: 300 messages a day on Claude Sonnet 4.6, each about 3,000 input tokens (system prompt, recent history and retrieved knowledge) and 600 output tokens. A month of 9,000 messages costs $56.27, against $162 at official prices. The same month on GPT-6 Luna would be $0.804.
• Routing workflow: a Question Classifier node on Claude Haiku 4.5 sorts 1,000 requests a day, about 800 input and 300 output tokens each. A month of 30,000 calls costs $24.06, against $69.00 at official prices.

Knowledge retrieval makes each request bigger, so the number of chunks you retrieve matters as much as the model. Your Dify logs and the APIVAI Dashboard both show tokens per request. For a wider price comparison, see Claude API pricing in 2026.

Only for the answering part. A Dify knowledge base needs an Embedding Model to index and search documents, and APIVAI does not provide embeddings. Install a separate embedding provider and set it under Default Models → Embedding Model (a Rerank Model is optional and also comes from another provider). The LLM that writes the answer from the retrieved text can still be an APIVAI model.

• Credentials validation fails or 401: check that the key was copied in full, still exists and has balance or budget left.
• 404: the /v1 is wrong. API Base URL in OpenAI-API-compatible must be https://api.apivai.com/v1, with /v1 once. API URL in the Anthropic plugin must be https://api.apivai.com; with /v1 the requests go to /v1/v1/messages.
• Model not found: Model Name must be the exact ID, such as claude-sonnet-4-6, not a display name. Check it against /v1/models.
• Empty or cut-off answers: raise Max Tokens in the app to 4096 or more, and raise Upper bound for max tokens in the model entry if the slider stops too low.
• The agent never calls tools: edit the model entry and set Function Call Type to Tool Call.
• Self-hosted Dify cannot connect: the Dify containers that run plugins need outbound internet access to api.apivai.com.
• 429 Too Many Requests: the default limit is 60 requests per minute per key. Batch workflows can hit it; slow them down or ask support for a higher limit.

No. One OpenAI-API-compatible provider holds all your models; you add one entry per model ID with the same key and API Base URL.

Yes. Both use the same provider plugins and fields. A self-hosted instance needs outbound access to api.apivai.com.

Yes. Each LLM node chooses its own model, and one APIVAI key covers every model.

No. APIVAI does not offer embeddings, so Dify Knowledge needs a separate embedding provider. Chat, agents and workflow LLM nodes work through APIVAI.

Request and response content is not logged. Only usage data such as model, tokens and cost is kept for billing. Keys and limits are explained in the docs. If you also automate with n8n, see n8n with Claude and GPT.

Create an account, top up from $10 and paste your key into Dify.