China National Supercomputing Center
SCNet Coding Plan — Lite / Pro tiers backed by China's national 100K-card compute pool
Why we recommend China National Supercomputing Center
Why we recommend SCNet
The National Supercomputing Network (SCNet / 超算互联网) is China's national-grade compute backbone, purpose-built for AI and HPC workloads. The core node went live in 2026 with the country's largest domestic AI compute pool activated, and its storage system topped the IO500 rankings. The Coding Plan is SCNet's developer-facing subscription product under its Model Service umbrella — packaging "domestic AI compute + mainstream coding tools" into two fixed-monthly tiers (Lite / Pro).
The Coding Plan design mirrors Zhipu / Qwen / Kimi — fixed monthly fee + soft quotas — but has a few notable distinctions:
- Domestic compute backbone: compute comes from SCNet's core node, emphasizing sovereignty and large-scale domestic AI accelerator support.
- Plan-dedicated API key:
sk-sp-prefixed Coding Plan keys are fully isolated from the generic API key (sk-prefix, pay-as-you-go). Mixing them up routes the call to pay-as-you-go billing and incurs additional charges. - Dual OpenAI / Anthropic Base URLs: both tiers expose
https://api.scnet.cn/api/llm/v1(OpenAI-compatible) andhttps://api.scnet.cn/api/llm/anthropic(Anthropic-compatible) — swapbaseURL+ key and existing code ports over. - Request count = model call count: a single user prompt typically triggers 5–30 model calls (5–15 for simple tasks, 15–30+ for complex ones); each call consumes one quota unit.
Plan scope and quota:
| Tier | Price | 5h quota | Weekly | Monthly | Best for |
|---|---|---|---|---|---|
| Lite (basic) | ¥20 / month | ~1,200 | ~9,000 | ~18,000 | Individual devs / light workloads |
| Pro (premium) | ¥100 / month | ~6,000 | ~45,000 | ~90,000 | High-frequency / complex projects |
Models currently available (refreshed as the platform adds more):
- MiniMax-M2.5 — general dev Q&A, code understanding, code modification, interactive programming.
- Qwen3-235B-A22B — code generation, engineering reasoning, complex instructions, multi-turn development collaboration.
Lite and Pro use the same models and inference service; the only difference is request quota. The Anthropic protocol currently supports MiniMax-M2.5 only — more models are being adapted. A free OpenClaw 2-core / 4-GB instance is bundled with every subscription.
The signup flow is straightforward: log in to scnet.cn → Console → Model Service → Coding Plan → pick a tier and pay → grab the sk-sp- key and Base URL on the same page → paste into Cursor / Claude Code / Cline / etc. Quota exhaustion does NOT auto-fall back to pay-as-you-go (pay-as-you-go uses the generic API key instead). The subscription period extends from purchase time to the same day next month at 23:59:59 UTC+8.
If you need domestic Chinese compute + a fixed-monthly Coding Plan + dual OpenAI / Anthropic protocol access, SCNet Coding Plan is a strong 2026 pick in the China-based subscription tier landscape.