If you code with AI all day — agents, documentation, batch jobs — the bill adds up fast. Pay-as-you-go API pricing is wonderfully flexible right up until you see the invoice. For most of my day-to-day work, I’ve switched to subscription-based coding plans, mostly from Chinese providers, which trade a little flexibility for a much friendlier price. ROI matters, no matter if you are paying for it yourself or by your company, developers are disproportionately rewarded for their optimised solutions.
A quick bit of context on why I care about this. I split my time between Hong Kong and Shenzhen, which turns out to be a near-ideal setup. In Hong Kong I pay just HK$199 a month (about US$25) for symmetric 5 Gbps fibre — that’s what most of my AI-agent web research runs on, since all that crawling and fetching eats bandwidth both ways. Across the border in Shenzhen, I pay ¥159 a month for 2500/150 Mbps, plus something the HK side can’t easily give me: frictionless access to the local Chinese model providers, where the Alipay/WeChat payments and local endpoints that trip up overseas users just work. That combination is how I keep costs near zero while building odw.ai and the suite of ten products we’re launching — in the open, with the source shared for anyone who wants it.
How to choose a coding plan?
Comparing these plans by hand is miserable — every provider prices differently, and the tiers rarely line up. I found these coding plan comparison sites recently:
| If you want… | Best site |
|---|---|
| Chinese-vendor coding plans, side-by-side dashboard | codingplan.org |
| Same, but broadest coverage incl. long-tail vendors (Huawei, JD, Tencent, StepFun, Xiaomi…) | z4crk6mg95.coze.site + jia.je |
| Raw normalized data you can fork | github.com/mahonzhan/awesome-coding-plan |
| Global (Cursor / Copilot / Claude / Windsurf) with live pricing | pricepertoken.com/coding-assistants |
| Biggest directory of agents (30+) | artificialanalysis.ai/agents/coding |
| Pairwise deep-dives (X vs Y) | kilo.ai/compare |
| Broadest vibe-coding tool catalog | vibecoding.app/tools |
| Performance / speed benchmarks (not price) | coding.15o.cc + morphllm.com/ai-coding-agent |
| Editorial buying-guide, Chinese-market | Zhihu 完整指南 |
| Team/enterprise-tier detail (¥/seat) | AtomGit report |
— note: some activation links appear to route through referral domains, I am not related to any of them.
Why coding plans, and when
The appeal is simple: used well, they’re far cheaper than metered API access. The trade-offs are two. You’re limited by usage pattern — you can’t blitz your whole allowance at once, since caps usually apply to a rolling window (most often any 5-hour period). And you’re limited by models — cheaper tiers lock you out of the strongest models or burn quota faster when you use them.
So they’re my go-to for anything non-critical and non-urgent. If a task can wait or be chunked, it’s a great fit; if I need a guaranteed answer right now, I pay for on-demand API and move on. The 5-hour window sounds restrictive but rarely is, because agents run unattended: queue a backlog and drip-feed it, batch the heavy jobs overnight, and use a cheap model for the boring 80% while saving the pricey one for the parts that need it.
Can you use them outside China?
Mostly yes — but several providers run two platforms: a mainland version (CNY, Alipay/WeChat) and an international one (USD, foreign cards, a different endpoint). The comparison site reflects this, listing domestic and “international” versions separately.
The most foreigner-friendly are Zhipu/GLM (via its global site z.ai) and MiniMax (USD pricing, dedicated international endpoint) — both built to work from abroad. Just sign up for the right version and point your tools at the international endpoint, or they’ll silently fail. DeepSeek is pay-as-you-go and works fine abroad; the usual snag is your own bank blocking a China merchant, which a quick call fixes. The more domestic players — iFlytek, Volcano Ark, Xiaomi MiMo, Bailian — may want a Chinese phone number or ID and a Chinese-only console.
Two things regardless of provider: payment is the main friction point (the USD platforms sidestep it), and some employers or jurisdictions restrict Chinese AI tools over data-residency rules — so able to access isn’t the same as allowed to use.
The bottom line
Coding plans won’t replace on-demand access for urgent, critical work, and truly uncapped token plans are still — let’s be honest — reserved for the rich. 😃 But for the growing pile of work that’s important yet not urgent, they’re an excellent, cheap fallback. Work out your usage pattern, pick a plan whose quotas and models fit it, and let your agents handle the scheduling. It’s the same frugal approach that lets us build odw.ai and ship in the open — and the comparison site above is the best place to start.
Leave a Comment