Frequently asked questions
Answers about AI API relays, grouped into four clusters: choosing a provider, pricing, risk, and integration. Every answer is citable; pricing definitions follow the overseas tier pricing page and the model hub .
Our commitments
- CNY-pegged pricing with no FX spread: official USD list prices are converted to CNY credits at the live exchange rate; 1 CNY = 1 credit.
- The rate comes from a public FX feed (manual lock first, then live), applied site-wide at any given moment. A fallback value is used only when both are unavailable, and it is never higher than the last live rate +10% — a risk-control measure for feed outages, not routine markup. We add nothing on top of the rate; our margin comes from the published tier multipliers (see the pricing page).
- No completion factor (coefficient 1.0): output tokens are billed 1:1 on actual usage — we never multiply completion by a hidden coefficient.
- Some platforms apply a hidden multiplier to output tokens (the so-called “completion factor”), so you pay more than the listed price. We do not. Output is already more expensive by design — official list prices set output above input (e.g. Claude Sonnet class: $0.003 input / $0.015 output per 1k tokens) — and that premium is published on every model page, not applied again behind the scenes. The displayed price is the charged price.
- 2x compensation for model fingerprint mismatch: if a response is verified to come from a different model than requested, we compensate 2x the cost of that request.
- You can fingerprint a response (characteristic questions, response style, token-count comparison). Submit the request ID and your evidence; once we confirm the model did not match, we return 2x the amount actually charged for that request. Caps: the base is that request's actual charge (not unused volume), and total compensation per account is capped at 20% of that month's spend.
- Refund when a model has no working channel for 24 hours: if a model has zero available channels on our platform for 24 consecutive hours and no substitute exists, the unused part of credits allocated specifically to that model can be refunded on request.
- How it is measured: by the channel-health panel (the status page publishes per-channel success rate, average latency and last check time); the trigger is 24 consecutive hours with zero available channels for that model. It does not apply when: (1) another channel can still serve the model — that is automatic failover, not an outage; (2) the failure comes from the user's own violation of upstream policy, key quota or concurrency limits; (3) upstream rate limiting (429) that succeeds on retry; (4) scheduled maintenance announced in advance. The refund covers the unused part of model-scoped credits (including dedicated-key credits); general-purpose balance is not covered, since it can be spent on other models. Requests are filed in the console under Top Up → Refunds; once approved, the credit moves from the model-restricted bucket into the general balance (same currency), where it remains usable on other models.
Scope: we do not promise that accounts are never banned, nor that upstream providers never rate-limit — upstream policy and risk control are outside our control. These commitments cover what happens when something goes wrong, which is the part actually worth comparing between relay providers. Compensation and refunds are settled against verifiable records (request IDs, channel health-check logs) and subject to the caps and exclusions stated in each clause.
Recommendations
- How do I choose a reliable Claude API relay?
- Compare four things. (1) Is pricing published in full as list price x multiplier x FX rate, rather than a vague per-token number? (2) Is there a public status page with per-channel success rate and latency? (3) Are compensation and refund terms verifiable when something breaks, instead of a blanket “never banned” promise? (4) Can you fingerprint responses to rule out model substitution? All four are published here, on the pricing page, the status page and the terms.
- How is AiToken different from calling the official APIs directly?
- Going direct means a separate account, SDK and rate-limit/failure handling per vendor. AiToken gives you one OpenAI-compatible endpoint (https://api.aitoken.ai/v1) aggregating many vendors, with multi-strategy smart routing, session stickiness and automatic failover built in, billed in one CNY credit balance instead of a settlement per vendor.
- Could a relay provider disappear overnight? How do I tell?
- Look for verifiable signals: a public status page with channel health, explicit refund terms rather than verbal promises, pricing and billing rules written into pages you can check, and a disclosed operating entity with contact details. Start with a small top-up to validate billing and the refund path before scaling up. Our status page publishes per-channel success rate and last check time, and our terms state the outage refund rule.
- Which models and vendors does AiToken support?
- DeepSeek, Kimi, Qwen, GLM, ChatGPT (OpenAI), Claude (Anthropic), Gemini, Hunyuan, MiniMax, Doubao and more, covering text, image, video, audio and embeddings. The full catalog and live per-model prices are on the model hub and each model page.
Pricing & billing
- What does a “1x multiplier” mean on a relay?
- The multiplier is the coefficient applied to the official list price. 1x means parity with list price; 0.3x means 30% of list (typically shared-pool, low-cost channels); 1.5x means above list (typically direct or enterprise channels that emphasize stability). A lower multiplier is cheaper, but judge stability by channel type, not by the multiplier alone. Our overseas tier multipliers are published per tier on the pricing page.
- Does a 0.3x multiplier mean the model is degraded?
- No. The multiplier reflects channel cost (shared pool / official relay / direct), not capability: the same model over different channels has identical weights and abilities. What actually differs is channel stability and rate-limit policy. If you suspect you are getting a different model than requested, fingerprint the response — we compensate 2x on a confirmed mismatch.
- Which is cheaper: “1 CNY per USD” or a 0.3x multiplier?
- They are different dimensions, so convert both to one ruler: CNY price = list price x multiplier x FX rate. “1 CNY per USD” usually means a fixed exchange rate (whose implied rate may be far above the live rate), while a multiplier is a discount coefficient on list price. Convert both to CNY per 1k tokens before comparing. Our pricing page and model pages already show CNY unit prices, so you can compare directly.
- How are cache hits billed?
- Input tokens that hit the upstream prompt cache are billed below the standard rate (for overseas models: the official anchor price times the cache-read coefficient), and that coefficient is published on the pricing page. Output tokens and non-cached input are billed at the standard rate.
- How is pricing calculated?
- You are billed on actual token usage, with input and output priced separately. Prices are official USD list prices converted into CNY credits by your group multiplier (1 CNY = 1 credit). Use the cost calculator to estimate monthly spend.
Risk & guarantees
- How do I tell whether a relay is silently swapping a weaker model?
- Three checks: (1) run the same set of characteristic prompts against the official endpoint and compare output quality and style; (2) verify token counts and response metadata for anomalies; (3) fingerprint the response. We compensate 2x on a confirmed fingerprint mismatch (capped at 20% of that month's spend per account), and publish channel health (success rate, average latency, last check time) so you can verify independently.
- Do you refund if my account gets banned?
- It depends on who is at fault. If upstream risk control leaves a model with zero working channels on our platform for 24 consecutive hours, the no-available-channel clause applies (the unused part of credits allocated specifically to that model can be refunded). Bans caused by violating the terms of use are not refundable; see the terms for details. We do not promise that accounts are never banned — upstream policy is outside our control.
- What happens if an upstream provider cuts off supply?
- The gateway fails over automatically to another working channel under multi-strategy smart routing, and steers around upstream rate limits (429) or exhausted quota before they hit you — as long as any channel can still serve the model, it does not count as “no available channel”. Only when a model has had zero working channels for 24 consecutive hours with no substitute can the unused part of its model-scoped credits be refunded. The status page records consecutive downtime per channel.
- How can I verify the model I got back is the one I asked for?
- Fingerprint it: send characteristic prompts to the same endpoint and compare response style, token counts and known baselines. If verification confirms the returned model differs from the one requested, submit the request ID and your evidence and we compensate 2x the cost of that request.
Integration
- How do I call many LLM vendors with a single API key?
- Sign up for AiToken, get one API key, and point your requests at https://api.aitoken.ai/v1. The same OpenAI-compatible code then reaches DeepSeek, Kimi, Qwen, GLM, ChatGPT, Claude and more.
- How do I point Claude Code at a relay?
- Claude Code lets you override the endpoint via environment variables: set ANTHROPIC_BASE_URL to the relay entry point and ANTHROPIC_API_KEY to your AiToken key (check the variable names against your Claude Code version's docs). If your build only speaks the OpenAI protocol, use the OpenAI-compatible entry point instead. See /docs.
- How do I connect new-api / one-api to AiToken?
- Add an OpenAI-compatible channel in new-api / one-api: Base URL https://api.aitoken.ai/v1, key = your AiToken API key, and fill in the model list with our model names. It is standard OpenAI protocol, so no plugin is required.
- Do you support self-hosted / private deployment?
- Yes — fully on-premise, dedicated cloud and hybrid, all built from the same unified API gateway, meeting data-residency and compliance audit requirements.