A night-time discount on Alibaba Cloud Bailian: qwen3.8 models started after 22:00 are billed at 40% (3.8-Flash lands near ¥0.32/¥1.08 per million tokens). It is an official billing promotion rather than resale pricing; coverage and model lists follow the campaign page.
A free API model for daily Q&A, a supercomputing token bundle for batch runs, and a monthly coding plan to cap heavy usage — three tiers that keep a personal AI budget adjustable from zero to about ¥100 a month, with a fallback whenever one promotion ends.
A domestic aggregation API with DeepSeek, Qwen, Doubao, Kimi and Step models on the shelf. DeepSeek-V4.1-Flash is priced in line with the official rate, some lightweight models run as low as ¥0.7/¥2.1, and a batch of small models is free. Good for prototyping and vendor fallback; keep core production on official endpoints.
A monthly coding plan from Zhipu: the Lite tier is ¥94.4 on auto-renew (118 list price at 20% off) and lets you use GLM-5.3 and GLM-5.3-Flash inside Claude Code, Cursor, OpenCode and 20-plus other tools without token billing. Quotas are capped per 5 hours and per week, and apply only to supported tools.
An overseas model aggregator whose public list currently shows 460 entries; switching models means changing one name, with usage broken down per model. Prepaid credits mean no arrears, and 16 zero-price entries let you test first. International payment is required, and aggregated prices may differ from vendors’ own.
DeepSeek’s official off-peak pricing: ¥1 input and ¥4 output per million tokens in off-peak hours, double at peak, with cache hits as low as ¥0.02. A 1M context, 384K max output and 2500 concurrency suit batch jobs whose schedule you can shift.
A lightweight model on the MiniMax open platform with a one-million-token context and text, image and video input. Official pricing is ¥1 input and ¥4 output per million tokens, ¥0.10 on cache hits — built for whole-document Q&A rather than heavy programming.
An open-source text model free to call on Zhipu’s platform: 30B class, 200K context, up to 128K output tokens, with function calling and MCP. Suited to prototyping and low-volume use; the free tier has concurrency limits.
An open-source translation model from Tencent Hunyuan covering 30-plus language pairs. It translates and does nothing else; usable straight from a model platform web UI, free for now, best suited to pure translation rather than reasoning.