The cheapest way to run a capable coding agent in mid-2026 isn't a coding tool at all — it's a backend. Alibaba Cloud's Model Studio Coding Plan is a flat monthly subscription that puts Qwen, GLM, Kimi and MiniMax behind an Anthropic-compatible endpoint. Point Claude Code at it and you get a working agent for roughly $50/mo instead of a metered API bill — with one clause you have to respect or you'll lose the key.
Here's the whole thing: what it is, how to wire it up, and where it bites.
What it is#
A fixed-fee subscription for interactive coding, sold by request quota rather than per token. The Pro tier is about $50/mo and includes roughly 90,000 requests a month; the cheaper Lite tier (~$10/mo) closed to new subscribers on March 20, 2026, so new users land on Pro. One subscription covers several model families — Qwen's own flagship plus GLM, Kimi and MiniMax — which is why it doubles as a cheap way to try open-weight coders without standing up your own serving.
The unlock for Claude Code users is that Alibaba exposes an Anthropic-compatible endpoint. Claude Code already lets you override its base URL and auth token, so nothing about the tool changes — only where the requests go.
Setup, start to finish#
- Subscribe to the Coding Plan (Pro) in Alibaba Cloud Model Studio and generate a Coding-Plan API key — it starts with
sk-sp-, which is how you know you're on the plan and not the pay-as-you-go API. - Grab the base URL from Alibaba's Claude Code setup page. The mainland and international regions use different hosts, so copy the exact one for your region rather than typing it from memory.
- Export the two variables Claude Code reads, then run it:
export ANTHROPIC_BASE_URL="https://<coding-plan-endpoint-from-alibaba>/apps/anthropic"
export ANTHROPIC_AUTH_TOKEN="sk-sp-..." # your Coding-Plan key
claudeThat's it — claude now drives a Qwen-class model through the plan, and your Anthropic account is untouched. (If you also want the GLM coder in Claude Code, the pattern is identical: a compatible endpoint plus a plan key.)
The tool is Claude Code. The model behind it is whatever endpoint you point it at. That indirection is the entire trick.
The clause that revokes your key#
Read this before you get clever: the Coding Plan is for interactive use inside coding tools only. Claude Code, Qwen Code, and similar interactive clients are fine. Wiring the sk-sp- key into an app backend, a CI job, a cron task, or any automated script is a terms violation, and Alibaba can suspend the subscription and revoke the key for it. The flat quota is subsidized on the assumption that a human is in the loop; automate it and you're outside the deal.
If you need an agent that runs unattended — overnight refactors, a backend service, parallel or scheduled runs — use metered API billing for that workload and keep the Coding Plan for hands-on work. They're different tools for different bottlenecks.
Two things to weigh first#
- Data residency. The endpoint is Alibaba-hosted. There's an international option, but both are operated by Alibaba Cloud — treat it like any external model provider and don't send secrets or regulated data you wouldn't hand to a third-party API. Check your own obligations before you route a client's codebase through it.
- Quality ceiling. A Qwen/GLM/Kimi-class model is strong and improving fast, but it won't match a frontier Claude model on the hardest reasoning-heavy tasks. This is a cheap daily driver, not a frontier replacement.
Who this is for#
If your bottleneck is budget, not model quality — a solo founder shipping high volumes of ordinary code, a small team doing a lot of non-frontier work — a ~$50 flat plan behind the agent tool you already know is one of the best price-to-capability trades on the board right now. It's the bottom of the coding-agent market that just split into a premium seat and a subscription floor: pick the end that matches what's actually slowing you down. For a full walk through the tiers, our solo-founder subscription guide maps price to bottleneck.
