Billing - OpenPAI Docs

OpenPAI uses a prepaid + token-based billing model, consistent with OpenAI. This page explains how the actual charge for each call is computed.

Billing components

Every successful call produces the following usage items:

Item	Meaning
`prompt_tokens`	Input tokens you send to the model (including system / messages / tools and everything else).
`completion_tokens`	Output tokens the model returns (visible content).
`reasoning_tokens`	Invisible chain-of-thought tokens from reasoning models (only o3 / GPT-5 / Claude thinking / Gemini thinking).
`cached_tokens`	Input tokens that hit the prompt cache.

Charging formula

charge =
    (prompt_tokens     × input price)
  + (completion_tokens × output price)
  + (reasoning_tokens  × output price)
  + (cached_tokens     × cache price)
  × user group multiplier

Where:

Input / output / cache prices come from each model’s official pricing, converted to RMB as USD price per million tokens × OpenPAI multiplier × exchange rate.
User group multiplier: determined by your account’s group; see Core concepts.
Cache price is typically 0.1× to 0.5× of the input price; see Cache billing.

Console log example

After a call, the Logs page shows token and billing details (without prompt / response content):

Field	Example
Model	`gpt-5.5`
Input / output / cache / reasoning tokens	`1024 / 512 / 800 / 0`
Multiplier	`1.0×`
Actual charge	`¥0.0034`
Request ID	`req_01ab...`

Billing by model type

Chat (Chat / Responses)

Charged by prompt_tokens + completion_tokens. If prompt caching is used, the cache-hit portion is settled at the cache price.

Reasoning (o3 / GPT-5 / Claude thinking)

reasoning_tokens are billed additionally at the output price (invisible but they consume compute). The console log breaks out thinking vs. answer tokens.

Embeddings

Charged by prompt_tokens (input) only; there is no output.

Image generation (DALL·E / gpt-image-1)

Charged by number of images + size, a fixed charge per call.

Audio (TTS / Whisper)

TTS is charged by input character count; Whisper is charged by audio duration (seconds).

Midjourney / Suno

Asynchronous task model, charged by number of tasks + action type. Failed tasks are not billed.

Rerank

Charged by the total tokens of (query + all documents).

Are failed requests billed?

Situation	Billed?
4xx client error (bad params, auth failure)	❌ Not billed
5xx upstream error that auto-retries successfully	✅ Billed for the successful attempt
All channels fail, client finally gets a 5xx	❌ Not billed
Streaming response disconnects midway	Billed by tokens actually generated
Client abort	Billed by tokens the upstream already generated

Out of balance

When account balance ≤ 0, all paid model calls immediately return insufficient_quota (HTTP 403). Some free models (such as Moderations) remain usable.

Check your spend

Console → Logs: per-call details.
Console → Usage statistics: spend charts aggregated by day / model / token.
API:
- GET /v1/dashboard/billing/usage — spend over a time range
- GET /v1/dashboard/billing/subscription — current balance
Response headers: every API response includes X-OpenPAI-Quota-Consumed and X-OpenPAI-Remaining-Quota.

Price changes

New models: launched at the official price converted at integration time.
Official price changes: OpenPAI follows the upstream, announced 24 hours in advance in the console.
Price cuts / promotions: pushed from time to time, subject to the console banner.

Quota & tokens

Account quota, token quota, expiry and transfers.

Cache billing

OpenAI / Claude / DeepSeek prompt-cache details.

​Billing components

​Charging formula

​Console log example

​Billing by model type

​Are failed requests billed?

​Out of balance

​Check your spend

​Price changes

​More