LLM API FAQ

Open the access page, choose a package, enter the delivery email, and complete payment. New credentials are delivered according to the checkout instructions; also check the spam folder.

Use the top-up page and enter the existing key carefully. Never paste a full key into a public ticket, group chat, screenshot, or shared document.

Check the payment status, delivery email spelling, inbox, and spam folder. If it is still missing after a short wait, contact support privately with the order number and payment time.

Use the official quota portal linked on this site. Enter the key only on that official page and never send the full credential to support.

Standard prepaid credits currently remain available until used. If a campaign, contract, or custom plan has different terms, the checkout page or signed agreement takes precedence.

Use https://gpt-agent.cc/v1 for North and South America, https://eu.gpt-agent.cc/v1 for Europe and Africa, https://hk.gpt-agent.cc/v1 for Southeast Asia, Hong Kong, Macao, Taiwan, Japan, Korea, Australia, and New Zealand, and https://api.llm-token.cn/v1 for mainland China.

A Base URL ends at /v1. A full endpoint adds /chat/completions, /responses, or /messages. Follow the field label in your client and do not append the same path twice.

Chat Completions-compatible, Responses-compatible, and Messages-compatible formats are available. Select the format expected by your library, client, or agent workflow.

Use the API Key, Token, or Authorization field. Raw HTTP requests normally send Authorization: Bearer YOUR_API_KEY. Remove quotes, placeholder text, and accidental whitespace.

Copy it from the live model catalog or pricing page. Display names may differ from request IDs, so preserve capitalization, hyphens, and version suffixes exactly.

The catalog covers major GPT, Claude, Grok, DeepSeek, Qwen, Doubao, Kimi, GLM, MiniMax, Hunyuan, StepFun, MiMo, and LongCat families. Check the pricing page for current versions, capabilities, and rates.

Match the capability first, then compare quality, latency, price, context size, and tool support. Run the same small test prompt across two or three candidates before committing production traffic.

Yes, for models marked with those capabilities in the live catalog. Confirm whether the selected model accepts image URLs, Base64 input, or produces images, and follow its dedicated request format.

Common setups include Codex CLI, Claude Code, OpenCode, OpenClaw, Cursor, Cline, Roo Code, Kilo Code, Windsurf, Cherry Studio, OpenWebUI, Dify, n8n, LangChain, and LM Studio. Use the tutorial for the exact field names.

Check the live catalog and service notices first. If only one model fails, temporarily switch to a comparable model; if several models fail, treat it as a routing or maintenance issue.

Credits are prepaid and deducted using the live rate for each model, including input/output tokens, multipliers, or per-task fees where applicable. Review the pricing page before purchase or production use.

Long conversation history, large outputs, automatic retries, agent tool loops, images, and cache behavior can all change usage. Keep prompts and retained context as small as the task allows.

Check for repeated history, long output limits, parallel requests, SDK retries, or agent loops. Compare the quota portal time range with application logs, without exposing the full API key.

Teams can contact support about volume, concurrency, procurement, invoicing, and support needs. Any custom pricing or service commitment must be confirmed privately in the applicable agreement.

Verify that the key is complete, has no whitespace, is placed in the correct field, and is used with one of this service's Base URLs. If exposure is possible, stop using and replace the key.

Check the exact model ID and endpoint path. Common causes are using a display name, duplicating /v1, or sending a Responses or Messages payload to a Chat Completions path.

First verify remaining credit, then lower concurrency and retry frequency. If balance is sufficient and 429 continues, keep the model, time, and error screenshot for support.

Try the endpoint closest to the deployment region, shorten context and output, and send a minimal request. If only one client fails, inspect its proxy, certificate, and timeout settings.

Retry once after a short pause and test another model. One-model failures usually point to that route; failures across several models may indicate shared routing or maintenance.

Send the same minimal request with the same key and Base URL, changing only the model ID. A single failure suggests a model route; several failures justify escalation to support.

Delete or retract the exposed message, stop using the key, and contact support for replacement or risk review. Mask all but a few leading and trailing characters in screenshots.

No. Public troubleshooting needs only the model name and an error screenshot. Account verification involving credentials must use a private, secure support channel.

Provide at most two items: (1) the model name and (2) an error screenshot that includes the time or request ID when available. Never include the full key, password, payment card, or unmasked identity data.

Check service notices first. After recovery, send one minimal request instead of enabling unlimited retries. If it still fails, report the model and the latest error screenshot.

Refunds and disputed usage follow the refund policy and actual order status. Stop further calls, then contact support privately with the order number and relevant time for review.