# How to call the 88z.ai model API: base URL, auth, limits, and how billing works

- Category: 行业应用
- Published: 2026-11-12T02:22:02.334Z
- Source date: 2026-11-12
- HTML: https://www.88z.ai/en/news/news-how-to-call-the-88z-ai-relay-api-20261112
- Markdown: https://www.88z.ai/en/news/news-how-to-call-the-88z-ai-relay-api-20261112.md
- Other language: https://www.88z.ai/news/news-how-to-call-the-88z-ai-relay-api-20261112.md

88z.ai offers an OpenAI-compatible subset API that aggregates models from major providers and bills in CNY points. This guide follows the live /api.md page: how to register, create a site key, point the Base URL at https://www.88z.ai/v1, which endpoints are implemented, what the limits are, and how charges are calculated. It is not the official Anthropic, OpenAI, or Google API, and it is not the same site as 88api.ai.

## At a glance

- Use https://www.88z.ai/v1 as the Base URL, with a key created on this site. Do not use a key from another platform.
- Implemented endpoints are the model list, non-streaming chat completions, and image generation; streaming, custom tool calls, embeddings, and audio or video output are not supported.
- Charges are billed in CNY points for actual usage (1 CNY = 1,000 points). Catalogue rates are a reference only; settlement uses the price version locked when the request was reserved.
- A request may include at most 32 messages and 64 KiB of text; set stream to false and send an Idempotency-Key.

## What it is, and what it is explicitly not

88z.ai, branded 88Z.ai (Ai最前沿), offers a web workbench and an OpenAI-compatible HTTP API that aggregates models from major providers. It is an AI model API platform with an OpenAI-compatible interface — one key calls Claude, GPT, Gemini, DeepSeek, Qwen and other major models — billed in CNY points. The canonical description is https://www.88z.ai/api.md, with an English edition at /en/api.md and a machine-readable catalog at https://www.88z.ai/models.json.

The same page draws several boundaries that should be read first. It is not the official Anthropic, OpenAI, or Google API. It is also not 88api.ai, which is a separate site. The canonical domain is www.88z.ai. The number of catalogue entries is whatever the live page shows that day, and only models that meet this site's text-chat rules can be called through /v1/chat/completions. The page does not provide latency measurements, and it does not claim that 88z.ai is necessarily the cheapest option or an official connection for any vendor.

## Register, create a site key, and point the Base URL here

Before you call the API, register at https://www.88z.ai/register with an email, a nickname, and a password of 12 to 256 characters, or sign in at https://www.88z.ai/login. Then open https://www.88z.ai/api and create a site key under API keys. The full key is shown only once, and the account needs enough points.

Set the client Base URL to https://www.88z.ai/v1. The documentation says the workbench page builds that URL from the current site origin, which is this address when you open the www host. https://api.88z.ai/v1/ points to the same interfaces, but the docs and examples default to https://www.88z.ai/v1. Authenticate with Authorization: Bearer YOUR_API_KEY, send Content-Type: application/json, and include Idempotency-Key with a unique request id you generate. A missing key returns 401 api_key_required; an invalid key returns 401 invalid_api_key. Use a key created on this site; do not use a key from another platform.

## Which endpoints exist today, and what the limits are

Three endpoints are implemented. GET /v1/models lists the text models currently available through this API; the full price catalog at https://www.88z.ai/models does not require a key. POST /v1/chat/completions accepts multi-turn system, user, and assistant text, and some models accept images, but only in non-streaming mode, so you must set stream to false. POST /v1/images/generations takes a model and a prompt, returns one image as base64 plus the charge, and the available image models are those shown in the workbench image picker.

The documented limits are at most 32 messages, 64 KiB of text, and 4 images per request. Images must be JPEG, PNG, or WebP data URLs, each at most 512 KiB after decoding, with an edge no longer than 2048 pixels, a total request body no larger than 4 MiB, and a model that actually accepts image input. The optional web_search flag is off by default. Streaming, custom tool calls, embeddings, and audio or video output are not supported. The server applies a reservation cap per request, and maxOutputTokens defaults to 4096. In clients such as Cursor, Claude Code, or Cherry Studio, use this site Base URL and key, the full model id, and turn streaming off.

## How billing works, plus a minimal request you can actually run

Billing reserves points up front from the request cap, then settles from that request's actual usage at the price version locked when the request was reserved. 1 CNY equals 1,000 points. Catalogue prices are a reference only, not a promised final invoice. Top-ups are denominated in yuan and designed for Alipay. Whether automatic credit is already enabled should be taken from the workbench and from the enabled field of GET /api/payments/config; this article does not treat that as already guaranteed.

A minimal non-streaming call looks like this. Prepare a site key and a unique Idempotency-Key, then POST JSON to https://www.88z.ai/v1/chat/completions with a model such as anthropic/claude-sonnet-5.5, one user message, max_completion_tokens set to 512, and stream set to false. If you use the official OpenAI Python SDK, set api_key to the site key, base_url to https://www.88z.ai/v1, and disable streaming; if the SDK enables streaming by default, you must turn it off explicitly. The SDK snippet in the docs follows the published HTTP contract and is not guaranteed there as a line-by-line production test, so a first integration should start with a small cap and confirm both the response shape and the charge.

## 88Z perspective: match the contract first, then wire it into your tools

The two easiest mistakes with a model API platform are to treat it as the official first-party API, and to assume it supports the full OpenAI feature set, especially streaming and tool calls. Reading /api.md rules both of those out quickly. Another common failure is the wrong key: you must use a key created on this site, not a key from another platform.

If you are connecting 88z.ai to an existing app, check three things first: that the Base URL and auth headers match the docs, that stream is false, and that the account has enough points for the reservation. The model catalog and prices change, so use the current contents of https://www.88z.ai/models and https://www.88z.ai/pricing.md. This article follows the public documentation; if the page is updated, the latest text wins.

## Sources

- [88z.ai API 说明](https://www.88z.ai/api.md) (2026-11-12)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
