Power Champion API Developer Guide

1. Overview

The Power Champion API is an OpenAI-compatible HTTPS interface. One key gives you chat, vision, embeddings, reranking, image generation, video generation, text to speech and speech recognition. Existing OpenAI SDKs work once you change the base_url and the key.

Scroll the table sideways to see every column.

Capabilities

Scroll the table sideways to see every column.

Conventions

• A missing required field returns 400 or 404. Optional fields fall back to their defaults when omitted.
• $PC_API_KEY in the examples stands for your key. Keep it in an environment variable, not in your code.
• Amounts are in US dollars. Billing endpoints also return exact values in nano-dollars (10⁻⁹ USD).
• All times are UTC. Daily quotas reset at 00:00 UTC (08:00 Taiwan time).
• Whether a model is online is defined by the service status. This guide does not guarantee that any particular model is available at all times.

2. Quick start

Three steps to your first request: get a key, set an environment variable, and send a request.

1. Get a key: email info@powerchampion.org to apply. Keys start with sk-b300- and are issued by our team and sent to you. The platform stores only a hash, so a key cannot be recovered — keep it safe.
2. Add credit: when you buy a credit pack you receive a redeem code. Apply it with POST /v1/redeem (see section 10).
3. Set the environment variable:

export PC_API_KEY="sk-b300-..."

curl

curl https://b300.powerchampion.ai/v1/chat/completions \
  -H "Authorization: Bearer $PC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.2-fp8",
    "messages": [{"role": "user", "content": "Describe Taipei in one sentence."}],
    "max_tokens": 512
  }'

Python

Install: pip install openai.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://b300.powerchampion.ai/v1",
    api_key=os.environ["PC_API_KEY"],
)

resp = client.chat.completions.create(
    model="glm-5.2-fp8",
    messages=[{"role": "user", "content": "Describe Taipei in one sentence."}],
    max_tokens=512,
)
print(resp.choices[0].message.content)
print(resp.usage)

Node.js

Install: npm install openai. Requires Node.js 18 or later, running as ESM (a .mjs file, or "type": "module" in package.json).

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://b300.powerchampion.ai/v1",
  apiKey: process.env.PC_API_KEY,
});

const resp = await client.chat.completions.create({
  model: "glm-5.2-fp8",
  messages: [{ role: "user", content: "Describe Taipei in one sentence." }],
  max_tokens: 512,
});
console.log(resp.choices[0].message.content);

Successful response

{
  "id": "chatcmpl-...",
  "object": "chat.completion",
  "created": 1792483207,
  "model": "glm-5.2-fp8",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "Taipei is the capital of Taiwan, …" },
      "finish_reason": "stop"
    }
  ],
  "usage": { "prompt_tokens": 14, "completion_tokens": 27, "total_tokens": 41 }
}

A 502 means the model is temporarily offline. Check the service status and switch to a model whose ready is true.

3. Authentication and key management

Every /v1/* and /dashboard/billing/* request must carry a Bearer key in its headers. No other authentication method is accepted.

Authorization: Bearer sk-b300-...

Scroll the table sideways to see every column.

Authentication failures

Scroll the table sideways to see every column.

Key security

• Use a separate key for each environment (development, testing, production) so you can disable one without affecting the others.
• Keep keys in environment variables or a secrets manager. Never commit them to version control or ship them in front-end code.
• If you suspect a key has leaked, email info@powerchampion.org right away to disable it and get a new one.

4. Model catalog and pricing

Nine models are available in production. Prices below are list prices on 2026-10-20. The authoritative source is the public catalog.json; if the two differ, catalog.json wins.

Language and vision models

Scroll the table sideways to see every column.

Retrieval models

Scroll the table sideways to see every column.

Generation and speech models

Scroll the table sideways to see every column.

“Status now” comes from the live status feed and refreshes about once a minute.

Video models

Scroll the table sideways to see every column.

Whether a video model is listed is defined by GET /v1/models (output_modalities contains video). Usage and billing are in section 9.

Availability

A model can go offline temporarily for maintenance or under load. Before calling one, check GET https://b300.powerchampion.ai/status.json (no key needed): only models whose ready is true are available. Calling GET /v1/models with your key returns the models that key can use and their per-token prices.

curl https://b300.powerchampion.ai/v1/models \
  -H "Authorization: Bearer $PC_API_KEY"

Each entry in /v1/models has id, context_length, max_output_tokens, pricing (US dollars per token, as strings), endpoint, input_modalities, output_modalities and supported_features. For non-text models, max_output_tokens is not the actual output limit.

5. Chat API

POST /v1/chat/completions uses the same request and response format as OpenAI Chat Completions, and supports streaming, tool calling, JSON output and image input.

Request parameters

Scroll the table sideways to see every column.

top_p, stop, seed and other OpenAI-compatible parameters are passed straight to the model, so their effect depends on the model. Parameters a model does not support may be ignored or return 400 — test before you go live.

Response fields

Scroll the table sideways to see every column.

Streaming

With stream: true, the response is text/event-stream. Each event is one data: {...} line, and the stream ends with data: [DONE].

• The platform turns on usage reporting, so any chunk may carry usage, and there may be a final chunk whose choices is an empty array. Check that the array is not empty before reading choices[0].
• If the backend fails after the stream has started, the connection simply closes without an error event. Tokens already generated are still billed.

stream = client.chat.completions.create(
    model="glm-5.2-fp8",
    messages=[{"role": "user", "content": "List three sights in Tainan"}],
    stream=True,
)
for chunk in stream:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

Tool calling

When you provide tools, the model may return tool_calls. Your code runs the tool and sends the result back as a role: "tool" message; the model then produces its final answer. In streaming mode, tool-call arguments arrive split across several chunks — concatenate them yourself.

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Look up the weather for a city",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}]

resp = client.chat.completions.create(
    model="glm-5.2-fp8",
    messages=[{"role": "user", "content": "What's the weather in Taichung right now?"}],
    tools=tools,
)
call = resp.choices[0].message.tool_calls[0]
print(call.function.name, call.function.arguments)

JSON output

response_format: {"type": "json_object"} asks the model to output valid JSON only. Describe the fields in your prompt as well. When you need strict fields, use json_schema instead (supported only by glm-5.2-fp8 and qwen3-vl-30b).

Reasoning models

glm-5.2-fp8 and glm-5.3-flash-uncensored support reasoning. The thinking process may use up part of max_tokens and may count toward output token usage. If a response has an empty content and finish_reason is length, raise max_tokens.

Image input

Use qwen3-vl-30b and put image_url entries in the content array, up to 4 images per request. Sending images as data: base64 is recommended.

import base64

with open("invoice.jpg", "rb") as f:
    b64 = base64.b64encode(f.read()).decode()

resp = client.chat.completions.create(
    model="qwen3-vl-30b",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "List the items and amounts on this invoice"},
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64}"}},
        ],
    }],
)

Personas

Power Champion can set up a fixed persona for you. Call GET /v1/personas/{persona_id} to get its messages_prefix and put it at the start of messages on every request. A fixed prefix hits the cache and lowers time to first token. Add ?rev=N to get a specific version; we provide the persona id.

{
  "persona_id": "...",
  "name": "Support assistant",
  "rev": 3,
  "updated": "...",
  "messages_prefix": [{ "role": "system", "content": "..." }]
}

6. Completions, embeddings and reranking APIs

Text completions POST /v1/completions

Legacy single-prompt completion in the OpenAI Completions format, with stream support. It works only with chat models; new projects should use the Chat API.

curl https://b300.powerchampion.ai/v1/completions \
  -H "Authorization: Bearer $PC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "glm-5.2-fp8", "prompt": "The highest mountain in Taiwan is", "max_tokens": 32}'

Text embeddings POST /v1/embeddings

Scroll the table sideways to see every column.

resp = client.embeddings.create(
    model="bge-m3",
    input=["Taipei Metro operating hours", "High-speed rail fares"],
    encoding_format="float",
)
vectors = [d.embedding for d in resp.data]  # 1,024 dimensions each

The response uses the OpenAI embeddings format: data[].embedding, data[].index and usage.prompt_tokens.

Document reranking POST /v1/rerank

Reorders documents by how relevant they are to a query — typically used to refine results after vector retrieval.

Scroll the table sideways to see every column.

curl https://b300.powerchampion.ai/v1/rerank \
  -H "Authorization: Bearer $PC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bge-reranker-v2-m3",
    "query": "How do I request a refund?",
    "documents": ["Refunds must be requested within 7 days", "Shipping is charged by weight", "Members earn points"],
    "top_n": 2
  }'

{
  "id": "rerank-...",
  "model": "bge-reranker-v2-m3",
  "results": [
    { "index": 0, "relevance_score": 0.9821 },
    { "index": 2, "relevance_score": 0.0137 }
  ],
  "usage": { "total_tokens": 46 }
}

results is sorted by relevance_score, highest first, and index points to the position in the original documents. Billing is based on the input tokens of the query and documents.

7. Image generation API

POST /v1/images/generations (also available as POST /v1/images) generates PNG images from text and returns them as base64. The endpoint accepts only the fields below; any other field returns 400.

Request parameters

Scroll the table sideways to see every column.

Example

import base64

resp = client.images.generate(
    model="flux-schnell",
    prompt="Jiufen Old Street in morning light, photorealistic",
    size="1024x1024",
    n=1,
    extra_body={"seed": 7},
)
with open("jiufen.png", "wb") as f:
    f.write(base64.b64decode(resp.data[0].b64_json))

Response

{
  "created": 1792483207,
  "data": [{ "b64_json": "iVBORw0KGgo...", "media_type": "image/png" }],
  "seed": 7,
  "steps": 4,
  "model": "flux-schnell",
  "guidance_scale": 0.0,
  "usage": { "prompt_tokens": 0, "completion_tokens": 1000, "total_tokens": 1000 }
}

Which model to choose

• flux-schnell: speed first — suited to large batches of drafts and real-time generation inside a product.
• chroma1-hd: quality first, with negative prompts and guidance_scale — suited to production assets.

Both cost US$0.01 per image. flux-schnell accepts negative_prompt, but it has almost no effect and makes generation slower; if you need a negative prompt, use chroma1-hd.

8. Text-to-speech and speech recognition APIs

Text to speech POST /v1/audio/speech

Turns text into speech, with emotion controlled by a text description. The response is the audio file itself. Streaming is not supported: the audio is returned once synthesis is complete, and the backend timeout is 300 seconds.

Scroll the table sideways to see every column.

Response headers: X-Audio-Seconds is the audio length in seconds, and X-Usage-Tokens is the billed amount (50 tokens per second).

curl https://b300.powerchampion.ai/v1/audio/speech \
  -H "Authorization: Bearer $PC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "indextts2",
    "input": "您好，歡迎來電，請問有什麼可以為您服務？",
    "voice": "<音色 id>",
    "instruct": "親切、溫和",
    "emo_alpha": 0.6
  }' \
  --output greeting.mp3

Split long text by sentence and synthesize each part separately. You wait less, and you stay under the 2,000-character limit.

Speech recognition POST /v1/audio/transcriptions

Transcribes audio to text. Chinese results are converted to Traditional Chinese (Taiwan). Upload the request as multipart/form-data.

Scroll the table sideways to see every column.

curl https://b300.powerchampion.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $PC_API_KEY" \
  -F file=@meeting.mp3 \
  -F model=whisper-large-v3 \
  -F language=zh

{
  "text": "今天的會議重點有三個…",
  "language": "zh",
  "duration": 42.6,
  "usage": { "prompt_tokens": 0, "completion_tokens": 2130, "total_tokens": 2130 }
}

Formats whose length can be read (such as wav and flac) return 400 when longer than 600 seconds; other formats are not guaranteed to be checked. Uploads over 25 MB return 413. Split long recordings into segments of 10 minutes or less before uploading.

9. Video generation API

POST /v1/videos generates an MP4 video from text, or from text plus one image. Video generation takes time, so it is an asynchronous job: submit the job, poll its status, then download the result. Authentication is the same as for every other endpoint: one key.

Available model

Scroll the table sideways to see every column.

Whether the model is listed is defined by GET /v1/models (output_modalities contains video). Resolution and frame rate are fixed and cannot be set in the request; a request that includes a field not in the table below (for example resolution) returns 400.

Flow

1. POST /v1/videos: submit the job. Returns 202 with the job id (status queued).
2. GET /v1/videos/{id}: poll the status, about every 5 seconds, until it is succeeded, failed, cancelled or expired.
3. GET /v1/videos/{id}/content: once succeeded, download the MP4.

Request parameters

Scroll the table sideways to see every column.

Example

curl https://b300.powerchampion.ai/v1/videos \
  -H "Authorization: Bearer $PC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan2.2-ti2v-5b-720p",
    "prompt": "A red fox walks through fresh snow at dawn",
    "seconds": 5
  }'

Success returns 202:

{
  "id": "0f3c…",
  "object": "video.job",
  "status": "queued",
  "model": "wan2.2-ti2v-5b-720p",
  "mode": "t2v",
  "seconds": 5,
  "estimated_cost_usd": "0.25",
  "expires_at": null,
  "created_at": 1760000000.0
}

The complete flow (submit, poll, download):

JOB=$(curl -s https://b300.powerchampion.ai/v1/videos \
  -H "Authorization: Bearer $PC_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"wan2.2-ti2v-5b-720p","prompt":"A red fox walks through fresh snow at dawn","seconds":5}' \
  | python3 -c 'import json,sys; print(json.load(sys.stdin)["id"])')

while :; do
  S=$(curl -s -H "Authorization: Bearer $PC_API_KEY" https://b300.powerchampion.ai/v1/videos/$JOB \
      | python3 -c 'import json,sys; print(json.load(sys.stdin)["status"])')
  echo "$S"; case "$S" in succeeded|failed|cancelled|expired) break;; esac; sleep 5
done

curl -s -o out.mp4 -H "Authorization: Bearer $PC_API_KEY" https://b300.powerchampion.ai/v1/videos/$JOB/content

Checking a job and job statuses

{
  "id": "0f3c…",
  "object": "video.job",
  "model": "wan2.2-ti2v-5b-720p",
  "status": "succeeded",
  "progress": 1.0,
  "mode": "t2v",
  "seconds": 5,
  "output_seconds": 5.041667,
  "cost_usd": "0.252090",
  "error": null,
  "content_url": "/v1/videos/0f3c…/content",
  "expires_at": 1760086400.0,
  "created_at": 1760000000.0,
  "finished_at": 1760000300.0
}

Scroll the table sideways to see every column.

expires_at and finished_at are Unix timestamps in seconds. About 7 days after a job expires its record is deleted and lookups return 404; until then the expired status can still be read.

Pricing and settlement

• US$0.05 per second of output video.
• Billed on the actual output length, capped at 1.1 times the requested seconds. The actual length can be slightly longer than requested (frames = seconds × fps + 1, so 5 seconds → 121 frames ≈ 5.04 seconds, about US$0.252).
• At submission, balance for the requested seconds × 1.1 is reserved (for example 5 seconds → 5.5 seconds = US$0.275). On completion the job is settled on the actual length and the difference is refunded.
• If a job fails, times out, or is cancelled while queued, the whole reservation is released and nothing is charged.
• The settled amount appears in cost_usd after completion; the estimated_cost_usd returned at submission is requested seconds × the unit price.

Download and cancel

• GET /v1/videos/{id}/content returns 200 with Content-Type: video/mp4, streamed. HTTP Range is supported: Range: bytes=… returns 206 (an invalid range returns 416), so downloads can resume.
• A finished video can be downloaded for at least 24 hours; the exact deadline is the job's expires_at. An unfinished job returns 409; an expired one returns 410 (404 once the record is deleted about 7 days later).
• DELETE /v1/videos/{id} cancels only queued jobs, with no charge, and returns {"id": "…", "status": "cancelled"}. A job that is processing or already finished returns 409; a job that has only just been submitted and is still being accepted can also return 409, so retry shortly. If the video backend is temporarily unreachable it returns 502 and the job is unchanged, so you can retry the cancel.

Errors on the video endpoints

Scroll the table sideways to see every column.

When a submission gets 429 or 503, no balance was deducted and you can simply submit again.

10. Billing: credit, balance and usage

The platform is prepaid: you add credit first, and each request is charged for its actual usage once it completes. When the balance runs out, requests return 402. Failed requests, such as backend or validation errors, are not charged, and neither is a non-streaming request whose connection drops before the response arrives. If a stream is interrupted after it has started sending, the tokens already generated are still billed.

Credit packs

Scroll the table sideways to see every column.

Credit packs are delivered as redeem codes. Email info@powerchampion.org to buy one.

Redeem a code POST /v1/redeem

curl https://b300.powerchampion.ai/v1/redeem \
  -H "Authorization: Bearer $PC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"code": "<redeem code>"}'

{ "ok": true, "credited_nano_usd": 10000000000, "new_balance_nano_usd": 12345000000 }

Scroll the table sideways to see every column.

Balance and usage

Scroll the table sideways to see every column.

curl https://b300.powerchampion.ai/dashboard/billing/subscription \
  -H "Authorization: Bearer $PC_API_KEY"

{
  "object": "billing_subscription",
  "has_payment_method": true,
  "hard_limit_usd": 8.42,
  "system_hard_limit_usd": 8.42,
  "access_until": null,
  "plan": { "title": "prepaid", "remaining_usd": 8.42 }
}

There is no API for per-request records yet; ask us for a monthly report. Every response also reports usage: usage for text models, and the X-Usage-Tokens header for text to speech.

Billing rules

Every model uses the same formula, priced at the list price when the request is accepted:

Cost (US$) = (input tokens × input price + output tokens × output price) ÷ 1,000,000

Non-text models are first converted into equivalent tokens:

Scroll the table sideways to see every column.

Video generation does not use the token conversion above: it is priced per second of output video, reserved at submission and settled on completion with the difference refunded. See section 9.

Precision is 10⁻⁹ USD. Charges are applied after the response completes, so the last request can leave the balance slightly negative.

11. Rate limits and quotas

Each key has four limits, and each model also has an overall capacity limit. The actual settings for your key are those agreed when it was issued — ask us if you need them. The platform does not send X-RateLimit-* headers.

Scroll the table sideways to see every column.

How the daily quota is reserved

When a request is accepted, max_tokens (2,000 if not set) is reserved from the day's quota, then corrected to actual usage when the request completes; failed requests release their reservation. When the quota is nearly used up, a smaller max_tokens lets more requests through.

Recommendations

• On a 429, wait for Retry-After and retry. If the daily quota is used up, wait until 00:00 UTC.
• For high-concurrency applications, add a queue in your code to keep in-flight requests within your key's limit.
• If you need higher limits, email info@powerchampion.org with your expected requests per minute, concurrency and daily usage.

12. Error handling

Errors use the OpenAI format. Only 429, 502 and 503 are worth retrying; for any other status code, fix the request first.

Error format

{
  "error": {
    "message": "Rate limit exceeded (60/min) for this key.",
    "type": "rate_limit_error",
    "param": null,
    "code": null
  }
}

Scroll the table sideways to see every column.

A few errors use {"detail": ...} instead: validation errors from the speech and image backends, 401 from /dashboard/billing/*, unknown paths (404), wrong methods (405) and query-parameter type errors (422). In addition, when the request body is not a JSON object (an array, for example), some endpoints may return a plain-text 500. Handle both formats when parsing errors. Error message text may change, so base your logic on the status code and type.

Status codes

Scroll the table sideways to see every column.

Check order

When a request has several problems at once, the platform checks them in the order below and returns the first failure. When debugging, work down the table from the top.

Scroll the table sideways to see every column.

Field-validation 400s (for example, an unsupported image field) and the speech-recognition 413 may be returned earlier than this order suggests.

Retry strategy

1. Retry only 429, 502 and 503, plus network-level connection errors.
2. If there is a Retry-After, wait that long. Otherwise use exponential backoff (1, 2, 4, 8 seconds) with random jitter, up to 4 attempts.
3. Use streaming for generations expected to take longer than 60 seconds; non-streaming requests longer than about 100 seconds may be disconnected. If a long non-streaming request times out, check your balance or usage to see whether it was charged before you resend it.

import random, time
from openai import APIStatusError

RETRYABLE = {429, 502, 503}

def with_retry(call, attempts=4):
    for i in range(attempts):
        try:
            return call()
        except APIStatusError as e:
            if e.status_code not in RETRYABLE or i == attempts - 1:
                raise
            wait = e.response.headers.get("retry-after")
            time.sleep(float(wait) if wait else 2 ** i + random.random())

resp = with_retry(lambda: client.chat.completions.create(
    model="glm-5.2-fp8",
    messages=[{"role": "user", "content": "Hello"}],
))

With the official OpenAI SDK, you can also set max_retries when you create the client and let the SDK handle retries.

13. Best practices and security

Performance

• Stream long answers: users see content right away, and you avoid proxy timeouts.
• Put fixed prefixes first: keep system prompts and personas unchanged and at the start of messages to hit the prefix cache and lower time to first token.
• Set a sensible max_tokens: it controls cost and reduces the daily-quota reservation.
• Retrieve with embeddings, then rerank: take the top 50 candidates with bge-m3, then the top 5 with bge-reranker-v2-m3 for the chat model.
• Split long audio: keep speech-recognition segments to 10 minutes or less, and split text to speech by sentence.

Reliability

• Have a fallback model for important flows: when the primary model returns 502 or 503, switch to a chat model that the service status shows as ready.
• Read status.json regularly for health checks instead of testing availability with real inference requests.
• Log each request's time (UTC), model, status code and usage, and include them when you report a problem.

Cost control

• Test with a small amount of credit before going live, and read /dashboard/billing/subscription regularly to watch the balance.
• Accumulate cost in your own system from each response's usage, and alert when the balance drops below a threshold.
• Choose flux-schnell when you only need drafts, and glm-5.3-flash-uncensored when you only need fast answers.

Security and compliance

• Keep keys on the server. Front-end applications should go through your own backend.
• Do not put passwords, keys or unnecessary personal data in prompts.
• glm-5.3-flash-uncensored has no built-in content filtering. Add your own content moderation to products for the general public.
• Have generated content reviewed by a person before publishing it, and label it as AI-generated where applicable law requires.
• For how data is handled and retained, see the Privacy Policy and the Data Retention Policy.

14. Appendix

A. Endpoint index

Scroll the table sideways to see every column.

B. Not currently supported

• OpenAI Responses API, Assistants, Files, Batches, Moderations.
• Audio translation /v1/audio/translations and realtime speech (Realtime / WebSocket).
• Single-model lookup /v1/models/{id}.
• Self-service key requests or revocation, online payment, a per-request usage API, and webhook notifications for request results.

C. Service status and support

Scroll the table sideways to see every column.

When reporting a problem, include the time (UTC), model id, endpoint, status code, the full error message, and the first 12 characters of your key (never the full key).

D. Changelog

Scroll the table sideways to see every column.