Getting Started With a Mainland-Accessible AI API Relay: DeepSeek/GLM/Kimi/Qwen Under One Key

📌 This guide solves three problems for mainland developers — registering at every model vendor separately, official prices running high, and multi-model management chaos — by sharing Tokeness, an aggregator I’ve tested with genuinely transparent billing. It fits coding, content creation, and batch processing.

🔗 Platform: tokeness.io

The three pain points

In 2026, using LLM APIs from mainland China means three recurring annoyances:

1. Registration sprawl: DeepSeek, Zhipu, Alibaba, Moonshot — every vendor registers, verifies, and tops up separately. Four models = four account systems.
2. High cost: official list prices have no discount, and with multiple models you pre-pay at every vendor.
3. Management chaos: keys, balances, and bills scattered across platforms; month-end reconciliation is painful.

Tokeness addresses all three with one account: full Chinese-model coverage, an OpenAI-compatible interface, CNY pay-as-you-go billing, and prices 13-44% below official.

Why this platform, not “just another relay”

Its differentiation is transparency and low friction:

• Full model coverage: DeepSeek V4 family, GLM-5.2, Kimi K2.6/K2.7 Code/K3, Qwen3.7 Max/Plus, MiMo V2.5, MiniMax M3 — one key for all
• Verifiable pricing: /api/pricing publishes input/output/cache multipliers per model; cache pricing isn’t hidden — rare in this industry
• Sane discounts: 56-87% of official, from real volume-purchase discounts — not below-cost-line bait
• Zero integration cost: OpenAI-compatible; change one base_url line; Cursor, Cherry Studio, and custom scripts all work
• Mainland direct access: no special networking required; register and use
• Security control: create multiple API tokens grouped by purpose; a single leaked token doesn’t expose everything
• Clear bills: every request returns full usage; backend keeps per-request records — you always know where the money went

Full registration + API key tutorial

3.1 Account registration

1. Open tokeness.io
2. Fill in registration: username, password, email, captcha
3. Log in to the console

3.2 Create an API token

1. In the console, find API Tokens → token management page
2. Click Add Token, give it a name (e.g. “Cursor dev”)
3. Optionally bind a model group to isolate permissions per tool
4. Create — you get an sk-xxxxxx key
5. Copy and save it immediately — the key is shown only once

Tip: create separate keys for testing and production so you can revoke one without touching the other.

3.3 Code integration

from openai import OpenAI

client = OpenAI(
    api_key="sk-your-token",
    base_url="https://n.tokeness.io/v1",
)

resp = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

Command-line verification:

curl https://n.tokeness.io/v1/chat/completions \
  -H "Authorization: Bearer sk-your-token" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"hi"}]}'

3.4 Tool configuration

• Cursor / Cherry Studio / Cline: in settings, set the API address to https://n.tokeness.io/v1, paste the token, and use the exact model ID from the plaza (e.g. glm-5.2)
• Custom scripts: use any OpenAI SDK; change base_url

Price reference

Verified 2026-08-01 (¥/M tokens):

The math: daily coding at 1M input + 0.3M output — about ¥2.9/day on deepseek-v4-pro, about ¥0.9/day on flash. Start with a small ¥20-50 top-up for a trial.

Scenario picks

• 💻 Development: Cursor codegen, error fixing, refactoring (deepseek-v4-pro / kimi-k2.7-code)
• 📝 Content creation: copywriting, logic outlining (qwen3.7-plus / qwen3.7-max)
• 🧠 Deep reasoning: long-document parsing, complex architecture (glm-5.2 / kimi-k3)
• 🔧 Batch processing: data cleaning, batch translation, automation scripts (deepseek-v4-flash / mimo-v2.5)

Common configuration issues

• 401 on calls: check the API key was copied fully, has no stray spaces, and isn’t disabled
• “Model not found”: the model name must exactly match the plaza ID (case-sensitive, version suffix included)
• Slow responses: rule out local network and proxy issues first; the same model can route through different upstream channels at different speeds — try another time window

Two pieces of advice

1. Start small: top up ¥20-50, run one day of real tasks, verify usage and billing are clean, then scale.
2. Don’t put sensitive data here: it’s a third-party aggregator — prompts pass through its servers. Personal code and public content are fine; customer data and production DB fields belong on official channels.

Platform: tokeness.io ｜ Prices verified 2026-08-01, subject to the live page.