AI API Relay Recommendations & Reviews (Updated 2026)

Before anything else

Updated 2026-08-01.

Using LLM APIs from mainland China is genuinely awkward: official platforms have annoying sign-up and payment flows and aren’t cheap. And picking a good relay is hard because you have to check all of these:

1. Stability, most important. A platform that’s constantly down or extremely slow costs you more time than it saves. Prefer long-running sites with public info and reachable support.
2. Speed. Slow responses are painful. Remember that the same model differs 5-10x across providers — test the specific model.
3. Model coverage. A good relay should call all mainstream models from one place, with fast new-model rollout.
4. Billing transparency. Many sites look cheap and run expensive. Token counting and rate multipliers are both traps. You need clear itemized bills, or you’ll never trace an anomalous charge.
5. Rug-pull risk. It exists at every site. Prefer company-style operations; watch support responsiveness and incident notices.
6. Cache pricing (the hidden one). The cache price should be ~10% of input price. Some sites charge 15%, bad ones 30%, and others don’t count cache at all. Coding workloads are cache-heavy — this single item can double the difference between two “same-price” sites.

Above all: this industry is unstable — don’t top up large amounts. Top up what you’ll use.

Prices change fast. I try to stay accurate, but trust each platform’s live rates.

Recommended

Tokeness

My current primary. The core reason is transparency: /api/pricing returns every model’s input, output, and cache multipliers directly — you can compute prices offline and don’t have to trust anyone’s word.

Model prices (verified 2026-08-01, ¥/M tokens):

Cache price: every model’s cache_ratio is published in the pricing API; cache price = input × cache_ratio. You can compute it yourself. In this industry that’s a rare move — too many platforms charge 30% for cache and don’t tell you.

Integration is standard OpenAI-compatible: base_url = https://n.tokeness.io/v1, one key for all models, works in Cursor, scripts, and Cherry Studio. Every request returns full usage; backend keeps per-request records.

Prices move (I personally watched flash rise 11% in a week) — but because they move, they track real upstream cost rather than being bait numbers.

New users: start with ¥20-50 and run your own real scenarios before deciding.

Neutral

Overseas aggregators (OpenRouter class)

The pioneers of the model: the most models, all original-vendor channels, decent stability. But foreign-currency cards are awkward now, prices run ~5% above official, and direct access from mainland China is unstable. Fine if you have overseas payment rails and need niche international models; otherwise not a great fit.

Official direct

DeepSeek, Zhipu, Alibaba, Moonshot official platforms are of course the most stable, with guaranteed model fidelity. The downside: separate registration and top-up at each, list price with no discount, and painful management/reconciliation when mixing models. If you use one or two models heavily, direct official access is entirely reasonable.

Open-source self-hosted (new-api / one-api class)

Deploy your own aggregation gateway — full data control, suitable for teams with ops capacity. But you own upstream channels, high availability, and billing yourself; hidden costs are significant. It fits as internal infrastructure for enterprises, not as an easy option for individuals.

Not recommended

Unknown relays from Xianyu/Xiaohongshu

I’ve tried several. It’s mostly heavily watered models or failing speed — pure luck. New sites are especially unreliable; I’ve watched more than one run off with balances. Be very careful.

Ultra-low-price sites (official 1-3x off)

A quote below the purchase cost line has exactly three explanations: reverse-engineered interfaces (banned at any moment, and their takedown can take your account too), billing black boxes (token padding), or a rug-pull. The official price is the cost line; anything far below it is a trap, not a bargain.

Per-request reverse proxies

Sites that charge per call instead of per token are usually heavy on padding and impossible to reconcile. If a site has a bad reputation, stay away.

The cache-price trap, in detail

This trap deserves its own section. All 2026 mainstream models support prompt caching: repeated prefix context is billed at the cache price, and the official cache price is ~10% of input. In coding, agent, and long-conversation scenarios, cache hits can be 50-80% of input.

The math: 100M input/month, 60M cache hits, input ¥10/M —

Same sticker price, 2.2x monthly difference. How to check: first, does the platform publish a cache price at all (Tokeness’s cache_ratio is the positive example)? Second, test it — send the same long prompt twice and check the second response’s usage for cache-hit detail and lower deduction. No cache detail = out for coding workloads.

Changelog

• 2026-08-01: initial release; recommended/neutral/not-recommended tiers + cache-price special section

Will keep updating: platform shutdowns, price changes, new platform entries. Star this repo.

Disclaimer

Review based on personal testing and public information; network conditions affect results. Prices follow each platform’s live pages. Don’t top up large amounts; top up what you’ll use — this applies to every platform, including my recommendation.