AIFast API FAQ

Clients that require the OpenAI Compatible Base URL usually fill in https://www.aifast.club/v1. If the client automatically appends /v1, the root domain name should be filled in based on the actual request log to avoid /v1/v1.

First carry Bearer Token to request /v1/models. When successful, a model list response containing a data array is usually returned, indicating that the basic authentication link is normal; returning 401 should check the Key and request header; returning HTML usually indicates that the address has not entered the API route.

Use the real model ID returned by the console or /v1/models. Do not use the web page title, Chinese alias or manufacturer name directly as the model parameter. Model supply and open status will change and are subject to real-time results.

Doesn't mean. Chat Completions, Responses API, Streaming Events, Tool Calls, Image and Video APIs are different capabilities. A successful basic chat can only prove that the current request is available, but cannot prove that the Agent or multi-modal functions are fully available.

It cannot be directly inferred. Claude Code uses Anthropic Messages semantics and requires the service to explicitly provide Anthropic-compatible entries and configure ANTHROPIC_BASE_URL and authentication variables according to official requirements.

The current Codex custom provider configuration uses the Responses API. In addition to base_url, env_key, and model ID, compatibility services also need to properly handle Responses API and tool call events; Chat Completions alone are usually not enough.

uncertain. Cursor official description says that custom keys are mainly used for standard chat models, and some special functions may continue to use Cursor’s own models. Settings may also vary by version and provider.

Vendors supporting OpenAI Compatible custom endpoints usually fill in https://www.aifast.club/v1. If the plug-in used will automatically append /v1, the root domain name should be filled in according to the final request log to avoid /v1/v1. Model IDs should be copied from /v1/models returned results.

First request /v1/models in the server or container where OpenWebUI is located, check the API Key, DNS, certificate and network connectivity; then check whether /v1 is repeatedly appended to the connection address. Just because the host is accessible does not mean that it is also accessible inside the container.

Connection tests may only verify domain names, certifications, or model lists. You should also confirm that the model selected in the session is consistent with the added model, that the model ID comes from /v1/models, that the final request goes to /v1/chat/completions, and that capabilities such as images and tool calls should be independently verified.

401 first checks Key, Bearer request headers and environment variables; 429 checks concurrency, frequency, balance and model busyness; 502 first uses short requests, turns off streaming output, and distinguishes upstream exceptions, timeouts and protocol incompatibilities.

Common reasons are that the model ID is misspelled, a display name is used, the model is not open to accounts, the API type does not match, or the client replaces the custom model name. Query /v1/models first and then check the request log.

Use environment variables or key management services, and don't write front-end code, Git repositories, screenshots, and public logs. Different devices, projects and environments use independent keys to facilitate statistics, quotas and revocation.

No. Model offerings, names, prices, context, and API capabilities are all subject to change. The tutorial only explains the verification method. The final information will be subject to the real-time return of the AIFast console, model price page and API.

At a minimum, test short requests, streaming output, timeouts, 429 backoffs, error logs, real model IDs, usage statistics, and key revocation. When Agent is involved, tool calls and file operations must be tested separately.

First use curl -N to directly request the origin site. If the terminal can display blocks one by one, it is usually caused by buffering in the application framework, reverse proxy, CDN or front-end reading logic; if the origin site also returns all at once, you should confirm whether the target model and API support streaming output.

Input usage, output usage, failed retries, and all requests generated by the Agent tool loop should be counted, rather than just the final successful one. The unit price and magnification of the model are subject to the real-time display on the price page and console.

There is no fixed number of times that applies to all businesses. Retry-After should be read (if present), use exponential backoff with random jitter and set the maximum number; configuration errors such as 400, 401 and model not found should not be automatically retried.

Use a temporary limited key to verify the live model list, a minimal text request, the response model field, usage, streaming, and required business capabilities. Then check rate limits, billing details, key management, and migration options. Do not decide from homepage model names, one speed test, or the lowest unit price alone.

Except for the Base URL and API Key, keep the model ID, prompt words, temperature, maximum output length, timeout, and question set consistent; repeat the test in a similar period, and record the success rate, P50/P95 delay, token, key capabilities, and complete task cost respectively.

cannot. Black box detection can only check whether the protocol, model declaration and behavior sample are consistent, and cannot prove the identity of the underlying model based on a single result. Comprehensive judgment should be made based on service provider disclosure, request records, multi-period retesting and real business baselines.

Not recommended. Temporary, low-limit, independent Keys that can be revoked at any time should be created and rotated or deleted immediately after testing; do not submit production keys, user data, internal prompt words, and undesensitized files.