Price of a Milk Tea, Buy a Monthly Subscription for the Top Four Domestic Flagship Models

9.9 yuan.

Not a promotional price from a small company, not a limited-time trial coupon, not a "first month special then full price renewal" word game.

It's the Token Plan monthly price from the National Supercomputing Internet — 9.9 yuan/month, up to 80 million tokens, GLM-5.2, DeepSeek V4, MiniMax M3, Kimi K2.6 freely switchable.

When was the last time something you bought for 9.9 yuan could help you write code, fix bugs, refactor architecture, and summarize million-word documents?

Why the National Team is Selling Tokens

Let's start with the background.

The National Supercomputing Internet is a national-level computing platform initiated in 2023 under the guidance of the Ministry of Science and Technology. It's not a startup, not a reseller, but a bona fide "national team."

What it does is simple: integrate the country's supercomputing and intelligent computing resources and provide services via the internet. In the past, to use supercomputing, you had to submit applications, queue, and wait for approval. Now, you register an account, top up, call API — just like using a cloud service.

But the "national team" doing this has a fundamental difference from commercial platforms: they don't rely on selling tokens to make money.

The positioning of the Supercomputing Internet is "computing affordability" — lowering the barrier for AI R&D so that university faculty and students, grassroots developers, and small innovative teams can also access top-tier computing power. The 9.9 yuan Token Plan price is essentially a subsidized price. Commercial platforms sell tokens as a business; the Supercomputing Internet sells tokens as infrastructure.

This determines two things: prices will remain low for the long term, and stability is backed by national-level computing resources.

What Can 9.9 Yuan Actually Do

The core logic of Token Plan: monthly credits are distributed, and input, output, and cache are all unified deductions. It's not per-call billing, but usage-based deduction.

During the promotion, 9.9 yuan/month can get up to 80 million tokens. Newly registered accounts can also stack tens of millions of free tokens.

What does 80 million tokens mean? Taking GLM-5.2 as an example, a complete code architecture analysis consumes about 50,000-80,000 tokens. 80 million tokens are enough for you to do 1,000 architecture analyses, or let AI help you write code for an entire month.

The key is that you are not model-bound. The same quota can be used freely with GLM-5.2, DeepSeek V4-Flash, MiniMax M2.7, and Kimi K2.5. No need to register, top up, and manage API keys separately for each.

Four Models, Each with Its Own Specialty

Zhipu GLM-5.2

Million-level context, code benchmarks comparable to high-end overseas models. If you need to handle ultra-long code files or do cross-file refactoring, GLM-5.2 is currently the domestic model with the largest context window.

DeepSeek V4

V4-Flash is fast and cheap, V4-Pro is more powerful. Stable performance in code completion, refactoring, and architecture sorting. DeepSeek's programming capabilities have been validated by many developers, so I don't need to say more.

MiniMax M2.7 / M3

Excellent deep thinking ability, suitable for complex business logic deduction. If you are doing requirement analysis, solution design — scenarios where you need to "think clearly before acting" — MiniMax is worth trying.

Kimi K2.5 / K2.6

Long text summarization and batch document parsing are strong points. Batch contract parsing, knowledge base Q&A, weekly and monthly report summaries — anyone who has used Kimi in these scenarios knows its efficiency.

Dual Protocol, One Key

This is the most hassle-free design of Token Plan.

The platform is compatible with both OpenAI and Anthropic protocol standards. A single key starting with sk-tp- works with both protocols.

The only difference is in the Base URL:

• OpenAI protocol: https://api.scnet.cn/api/llm/v1
• Anthropic protocol: https://api.scnet.cn/api/llm/anthropic

What does this mean? Your existing code, tools, and workflows can be used almost without modification.

Using Python OpenAI SDK? Just change one line of base_url:

from openai import OpenAI
client = OpenAI(
    api_key="sk-tp-your-key",
    base_url="https://api.scnet.cn/api/llm/v1"
)
res = client.chat.completions.create(
    model="glm-5.2",
    messages=[{"role":"user","content":"Analyze project call chain"}]
)

Using Claude Code? Set two environment variables:

export ANTHROPIC_BASE_URL=https://api.scnet.cn/api/llm/anthropic
export ANTHROPIC_API_KEY=sk-tp-your-key

Using Cursor? Fill in the Base address and key in settings, pass verification, and it's ready.

ccgui, OpenCode, Codex CLI — any tool that supports custom API addresses can connect.

Three Billing Plans, Don't Mix Them

The platform has three plans, with completely independent key formats and billing logic:

Mixing them will cause problems — quotas cannot be deducted, and extra charges apply. If you choose Token Plan, always use the key starting with sk-tp-. Don't use sk- keys to call.

Not Just API, but a Complete R&D Toolchain

Behind Token Plan is not just model calls.

The Supercomputing Internet has built a complete AI R&D toolchain: data preprocessing, model training, parameter fine-tuning, simulation computing, and inference deployment — full coverage. Notebook inference, small-sample fine-tuning experiments can all use Token Plan quotas.

There is also low-cost OCR (1 cent per call), document parsing, image-text recognition, and bill/license extraction. For scientific research and office scenarios, you can get everything done in one stop.

Enterprise users have priority scheduling channels, with shorter waiting times during peak hours. Complete usage ledger records every input and output token consumption, making team cost accounting and audits convenient.

Honestly Speaking, Limitations

The 9.9 yuan price is a promotional price. The original price is 50 yuan/month. When the promotion ends, the official haven't clearly stated.

80 million tokens is "up to" not "fixed every month." The specific quota depends on activity rules and stacked benefits.

The platform's model library is continuously updated, but some models' inference speed and stability may differ from the manufacturers' native platforms — after all, there is an extra layer of scheduling.

Also: Token Plan quotas are distributed monthly, unused portions do not accumulate. Don't hoard; use as needed.

Registration and Access

1. Visit Supercomputing Internet: https://www.scnet.cn
2. Register an account and complete identity verification
3. Enter the model service section, select the Token Plan promotion package, and subscribe
4. Obtain the exclusive API Key starting with sk-tp-
5. Configure the Base URL and key in your tools, then start using

Newly registered users can also receive tens of millions of free tokens. During the promotion, invite friends to complete identity verification to unlock up to 5,000 card-hours of computing power.

9.9 yuan can't buy you much harm, but it can buy you an AI programming monthly card on a national-level computing base.

This is not a startup's fundraising story; it's infrastructure guided by the Ministry of Science and Technology lowering the barrier. The difference: fundraising stories talk about "free in the future," while infrastructure is already "cheap now."

Project URL: https://www.scnet.cn