The platform's only unit is credits; 60 credits = ¥1 (machine-readable at GET https://zhonkezhonkeapi.dflop.top/api/v1/config/currency). Billing basis by modality: Text models: per input / output token, with a discounted rate for prompt-cache hits Image models: per generated image Video models: per second of finished video Speech synthesis: per input character; voice cloning and avatar creation per call Each call reserves an upper bound up front and settles against actual usage afterwards; the difference is returned. Per-model unit prices are listed in the model catalog.

FAQ — API docs

Q: What does an API key look like, and where do I create one?

sk-gpushare- followed by 64 hex characters (76 characters total). Create one in the console under Dashboard → API Keys → Create Key; the raw key stays retrievable from its detail page (stored encrypted). Every key can carry its own model allowlist, expiry, on/off switch, rate and concurrency caps, and spend limit — for isolation and audit. See Authentication.

Q: Which authentication methods are supported?

Four, pick one: the x-api-key header (recommended), Authorization: Bearer, the x-goog-api-key header that the Gemini SDK sends by default, or a ?key= query parameter. All three SDKs work with their default auth behaviour.

Q: Which plan should I pick?

Free ¥0: 100 credits per daily login, selected free models, creation platform Plus ¥168/month (¥1,680/year): 1,400 credits per week, all models unlocked Pro ¥888/month (¥8,880/year): 7,000 credits per week, higher rate and concurrency Enterprise SaaS from ¥9,800: a dedicated org space with a shared credit pool Private deployment from ¥98,000: the whole platform inside your own network Credit top-up packs (¥100 = 5,000 credits, with bonuses at larger sizes) are available at any time without changing plan.

Q: Is streaming supported?

Yes, on all three protocols: pass stream=true on the OpenAI side, use messages.stream() on the Anthropic side, and streamGenerateContent on the Gemini side. See Streaming.

The questions we get asked most, each with an answer you can act on directly. For full parameters and schemas, start from the API Reference.

What is Model Plaza?#

Model Plaza (zhonkemodel.dflop.top) is a multi-protocol LLM API gateway and an AI creation platform. One sk-gpushare- key calls models from Anthropic, OpenAI, Google, xAI, DeepSeek, ByteDance, Moonshot, Alibaba, MiniMax, Zhipu and Tencent, covering chat, image generation, video generation, music, speech, digital humans and knowledge-base retrieval. The full catalog with live prices is at the model catalog.

Do I have to change my code?#

No. The gateway speaks the OpenAI, Anthropic and Google Gemini wire protocols, so you only change two lines — base_url and api_key:

OpenAI SDK: base_url="https://zhonkezhonkeapi.dflop.top/v1"
Anthropic SDK: base_url="https://zhonkezhonkeapi.dflop.top" (no /v1)
Gemini SDK: http_options={"base_url": "https://zhonkezhonkeapi.dflop.top"}

Your SDK and the model you call are independent: you can call Claude with the OpenAI SDK, or GPT with the Anthropic SDK. Protocol translation happens inside the gateway. See the Quickstart.

What does an API key look like, and where do I create one?#

sk-gpushare- followed by 64 hex characters (76 characters total). Create one in the console under Dashboard → API Keys → Create Key; the raw key stays retrievable from its detail page (stored encrypted). Every key can carry its own model allowlist, expiry, on/off switch, rate and concurrency caps, and spend limit — for isolation and audit. See Authentication.

Which authentication methods are supported?#

Four, pick one: the x-api-key header (recommended), Authorization: Bearer, the x-goog-api-key header that the Gemini SDK sends by default, or a ?key= query parameter. All three SDKs work with their default auth behaviour.

How is usage billed?#

The platform's only unit is credits; 60 credits = ¥1 (machine-readable at GET https://zhonkezhonkeapi.dflop.top/api/v1/config/currency). Billing basis by modality:

Text models: per input / output token, with a discounted rate for prompt-cache hits
Image models: per generated image
Video models: per second of finished video
Speech synthesis: per input character; voice cloning and avatar creation per call

Each call reserves an upper bound up front and settles against actual usage afterwards; the difference is returned. Per-model unit prices are listed in the model catalog.

Is there a free allowance?#

Yes. New accounts get 120 credits, enough to run every example in these docs, and free-tier users get another 100 credits per day on login plus unlimited use of the selected free models. See Pricing.

Which plan should I pick?#

Free ¥0: 100 credits per daily login, selected free models, creation platform
Plus ¥168/month (¥1,680/year): 1,400 credits per week, all models unlocked
Pro ¥888/month (¥8,880/year): 7,000 credits per week, higher rate and concurrency
Enterprise SaaS from ¥9,800: a dedicated org space with a shared credit pool
Private deployment from ¥98,000: the whole platform inside your own network

Credit top-up packs (¥100 = 5,000 credits, with bonuses at larger sizes) are available at any time without changing plan.

Yes. Every key under an account draws on the same account balance. Key-level settings are for permissions and audit, not budget isolation. To split accounting per project, use per-key spend limits and the per-key call log.

Is streaming supported?#

Yes, on all three protocols: pass stream=true on the OpenAI side, use messages.stream() on the Anthropic side, and streamGenerateContent on the Gemini side. See Streaming.

Can I use it from Claude Code, Cursor and similar clients?#

Yes, and usually by changing one environment variable or one base-URL field. Claude Code, Cursor, Cline, Continue.dev, Open WebUI, New API and FlopCode are all verified; step-by-step configuration is under Integrations.

How do I call image, video and digital-human models?#

Images go to /v1/images/generations and /v1/images/edits (billed per image). Video is asynchronous: submit a job, then poll it by ID (billed per second of output). Digital humans and smart clipping have their own endpoints. See Image / Video APIs and Digital Human APIs.

A call failed — what do I check first?#

The status code. 401 invalid_api_key means a wrong or disabled key; 402 quota_exceeded means an empty balance; 403 model_not_allowed means the key's allowlist excludes that model; 429 rate_limit_exceeded means you exceeded a rate or concurrency cap; 503 no_channel_available means no upstream is currently available for that model. Every error body carries a trace field — quote it and we can find that exact call. Full table in Error Codes.

What happens when an upstream provider fails?#

Each model is backed by several upstream channels ordered by priority. Retriable failures (connection errors, 5xx, 429, timeouts) fail over to the next channel automatically and the caller still gets a normal response. Explicit 4xx refusals (a content-policy rejection, for example) are not retried — another channel would only return the same refusal.

Is my data used for training?#

Not for training any model of ours. Requests and responses are used to serve the call, settle billing and populate your call log, which you can query per key in the console. Upstream vendors' own data policies apply to their side. See the Privacy Policy.

Is there an OpenAPI spec or machine-readable documentation for agents?#

Yes, a full set:

OpenAPI 3.1: https://zhonkezhonkemodel.dflop.top/openapi.json
RFC 9727 API catalog: https://zhonkezhonkemodel.dflop.top/.well-known/api-catalog
llms.txt: https://zhonkezhonkemodel.dflop.top/llms.txt (full text at llms-full.txt)
A2A Agent Card: https://zhonkezhonkemodel.dflop.top/.well-known/agent-card.json
MCP Server Card: https://zhonkezhonkemodel.dflop.top/.well-known/mcp/server-card.json
Agent Skills index: https://zhonkezhonkemodel.dflop.top/.well-known/agent-skills/index.json

Every doc page and every model page also has a Markdown twin — append .md to the URL, e.g. https://zhonkezhonkemodel.dflop.top/docs/quickstart.md.

Is there an MCP server?#

Two, for different jobs:

Knowledge-base retrieval: https://zhonkezhonkeapi.dflop.top/mcp (Streamable HTTP), authenticated with the same sk-gpushare- key, searching the knowledge bases you build on the platform. See Knowledge Base API & MCP.
Site content search: https://search.dflop.top/mcp, no authentication, searching every doc page and model page on this site — so an AI client can look up how to call a given model and what it costs. The same index is also reachable as REST: POST https://search.dflop.top/search with {"query": "..."}.

How is this different from using the vendors' own APIs?#

Three things: one key covers many vendors, so you skip per-vendor signup and top-up; the three SDK protocols are interchangeable, so switching models needs no code change; and the same account includes the whole creation platform (images, video, digital humans, infinite canvas) plus the Deski desktop agent. The trade-off is that pricing is set by the platform and is not identical to each vendor's list price.

What is Deski?#

Deski is the platform's desktop agent for macOS and Windows. It reads local files, edits code and runs commands on your own machine, opening a separate Git worktree per task so you review changes before they merge. It shares the account and balance with the gateway — see deski.dflop.top.

Can it be deployed privately?#

Yes. The enterprise edition installs the whole platform — gateway, creation workbench, knowledge base, asset library, agents, digital humans — on your own servers, so data never leaves your network. Accounts are provisioned by an administrator and funded from a company-wide compute package. See the enterprise section on the pricing page.