Skip to main content

​
Account

After email verification, new users automatically receive a trial credit; the exact amount is shown at sign-up. Registering via an invite link can stack an additional referral bonus.
Yes. Apply under Profile settings → Account security → Delete account. After deletion, remaining credit can be refunded within 30 days (minus a fee); it is forfeited afterward.
Sign-in page → Forgot password → receive a reset link by email.
Technically yes, but it’s not recommended. Create a separate token for each user and allocate quota in the console — this controls cost and enables call auditing.

​
Top-ups & invoices

It usually credits automatically within 1-2 minutes. If it’s still missing after 10 minutes, keep the order number / payment screenshot for later verification.
Yes.
  • Unspent balance within 30 days: refunded to the original method, minus payment-channel fees.
  • Over 30 days: no refund, but it can be transferred to another account or kept on the account balance for continued use.
  • Crypto / redemption codes: no cash refund.
Submit a request under Console → Profile settings → Invoices. Electronic general invoices are issued in 1-3 business days; special invoices require qualification info and take 5-10 business days. Only RMB orders can be invoiced.
Yes. When going through the business channel, follow the site announcements.

​
API calls

  • OpenAI-compatible protocol: https://openp.ai/v1
  • Claude / Gemini native protocols: https://openp.ai (without /v1; the SDK completes the path)
  • The token’s model allowlist doesn’t include this model → edit the token to add permission.
  • The model is only open to a higher tier → upgrade your account group.
  • The model is retired → confirm in the console model list whether it’s still available.
You hit the RPM / TPM rate limit. We suggest:
  1. Implement exponential-backoff retries on the client.
  2. Reduce concurrency or merge requests.
  3. Upgrade to a higher tier for a larger quota.
See Rate limits.
Partially, subject to console announcements. Most customers are better served by the Responses API instead.
Technically CORS is enabled, so you can. But it’s strongly discouraged to expose your API key in the frontend — forward through your own backend.

​
Models & billing

OpenPAI price = official published USD price × multiplier × exchange rate. Some models are slightly marked up due to supply-side costs (premium accounts / reserved pools); some domestic models are actually cheaper. The Console → Models page shows each model’s live multiplier.
The chain of thought for models like o3 / GPT-5 is invisible by design, but it counts toward usage and is billed. Claude thinking and Gemini thinkingConfig modes do return thinking content explicitly. See Reasoning & thinking.
It depends. When a repeated system prompt is long and fixed, the cache-hit portion is settled at 0.1-0.5× of the input price, typically saving 30-70% of input cost in RAG / agent scenarios. See Cache billing.
No. Async tasks that fail on the system side are fully refunded; tasks rejected by the upstream are billed by the compute already consumed.

​
Logs & deployment

No. Prompt content is not recorded; only token counts are tallied for billing.
Enterprise customers can get a self-hosted / on-premises deployment plan; the specifics are subject to site announcements.

​
Still have questions?

Error code reference

A complete reference of HTTP statuses and error codes.