① How to choose a model
| Task | Recommended | Why |
|---|---|---|
| Hardest reasoning (research / math / architecture) | DeepSeek GLM-5.3 / GLM-5.3 | Deepest chain of thought; GLM-5.3 carries a 1M context |
| Everyday coding (IDE agents / review) | DeepSeek-V4-Pro / DeepSeek V4 | Best balance of speed and quality |
| High-volume tasks (classification / extraction / batch) | DeepSeek-V4-Flash / DS DeepSeek-V4-Flash / Qwen Flash | Several times cheaper, with strong instruction following |
| Chinese-language workflows (content / long documents / localization) | Qwen 3.8 / Kimi-K3 / GLM | Strongest Chinese understanding |
| Vision / multimodal | Qwen 3-VL Plus | Image understanding at a price far below flagship vision models |
| AI coding / Qwen ecosystem | Qwen3.8-Max / 5.4 | Native support in agent frameworks that follow the OpenAI format |
Rule of thumb:Start real work at the mid tier (DeepSeek-V4-Pro / DS V4), move up to GLM-5.3 class only when the task is genuinely hard, and drop repetitive work to the Flash tier — the bill can be about 10× lower, with quality loss that is often zero.
② The three protocols
| Protocol | Endpoint | Best for |
|---|---|---|
| Chat Completions | /v1/chat/completions | OpenAI-compatible SDKs, Cherry Studio, most tools |
| DeepSeek Messages | /v1/messages | DeepSeek Code, DeepSeek SDKs, Cursor |
| Responses | /v1/responses | Coding tools and extensions, newer OpenAI-compatible SDKs |
③ How billing works
Each request is metered in four token classes: input (1×), output (about 5× input), cache reads (0.1× input — the biggest saving lever), and reasoning tokens (billed as output). Tip: keep your system prompt stable to maximize cache hits; streaming and non-streaming cost the same; console usage detail goes down to the token.
④ Why use a gateway instead of connecting directly?
⚡ One key, every model
No separate signup, separate top-up or separate dashboard per vendor. DeepSeek, Qwen, GLM, Kimi and other mainstream models share one balance.
💳 No overseas credit card
Top up by scanning a WeChat Pay or Alipay QR code, from ¥100, with no difference in treatment for small or large amounts.
🎯 Stable long-term pricing
Everything is billed at the official list prices published by each vendor — price is policy, not a promotion that expires.
🛡 Independent control per key
One key per project with its own limit, frozen or deleted from the console in one click.
⑤ Ways to save
- Keep the system prompt stable — cache reads cost one tenth of fresh input, the biggest saving lever.
- Start at the mid tier — DeepSeek-V4-Pro / DeepSeek V4 handles most real work at roughly a tenth of the flagship bill.
- Downgrade repetitive tasks — use the Flash / Flash tier for classification, extraction and batch processing, where quality loss is often zero.
- Read your usage detail — every console call is recorded at token level, so you can see exactly where the money goes.
⑥ Glossary
| Term | Plain meaning |
|---|---|
| API Key | Your credential key, sent in the request header Authorization: Bearer sk-sl-… |
| Base URL | The gateway address you put into a tool, such as https://api.shinloo.com/v1 |
| Context window | Maximum tokens a model can see in a single request (1M ≈ an entire codebase) |
| SSE / streaming | Token-by-token output, which makes chat and agent interfaces feel instant |
| Tool calling | The model asks your code to run a function — the basis of agents |
| Thinking / reasoning | A model that thinks before answering; stronger but more expensive per token |