Are the Cheap Chinese AI Coding Plans Legit?
Yes. The Chinese AI coding plans are legitimate and genuinely cheap: GLM Coding Plan (Z.ai) from $18/mo Lite, Kimi Code from $19, MiniMax M Plan from $22, all usable inside OpenCode, Cline, Claude Code harnesses and similar. Users report excellent value on routine coding ('pay 3 cents for $30 of work' is the flavor of the praise). The catches are real too: performance claims come from vendor-run benchmarks, every plan meters 5-hour and weekly windows, peak hours cost more or slow down, use is limited to interactive coding tools, and vendors have changed or closed plans mid-subscription (MiniMax closed its Token Plan to new buyers in favor of the M Plan).
The plans on the table
| Plan | Price | Strong at | The catch |
|---|---|---|---|
| GLM Coding Plan (Z.ai) | $18 / $80 / $168 per mo (Lite / Pro / Max) | High-volume routine coding | Credit quotas per 5 hours and per week; weekdays 14:00 to 18:00 Singapore time burn credits at full rate, other hours at half; supported coding tools only |
| Kimi Code (Moonshot) | $19 / $39 / $99 / $199 per mo (Plus to Ultra) | Long-context work | Rolling 5-hour window plus a monthly cap, amounts unpublished; interactive use only; K3 with 1M context needs Pro or higher |
| MiniMax M Plan | $22 / $55 / $132 per mo (Go / Explore / Build) | Text plus image, audio and video on one plan | 5-hour and weekly windows of unpublished size; replaced the Token Plan, which is closed to new buyers |
| DeepSeek via API | $0.22–0.66 per MTok input off-peak | Cheapest raw tokens anywhere | Peak/off-peak price doubling; API not a flat plan |
Plan prices and limits checked on each vendor's page on 2026-10-14; sources are linked in the flat-rate plan comparison.
Where they genuinely win
For routine implementation work (CRUD, refactors, tests, glue code), the current open-weight generation (GLM, Kimi, DeepSeek, Qwen families) is close enough to frontier quality that paying frontier prices is waste. At $18–22 flat, these plans are the cheapest real coding capacity on the market, and the community treats them as the standard budget layer: cheap plan for volume, frontier model for the hard 20%.
The catches, spelled out
- Vendor-run benchmarks: the '≈ Sonnet-class' claims come from the vendors; independent results are more mixed, especially on hard debugging and long agentic chains.
- Throughput, not just tokens: per-tier concurrency is mostly unpublished, peak hours cost more (Z.ai) or bring dynamic rate limits (MiniMax), and slowdowns are widely reported.
- Plan instability is the pattern, not the exception: quota semantics rewritten, cheap tiers killed, weekly caps added mid-2026 across several vendors.
- Data residency: your prompts and code go to infrastructure under Chinese jurisdiction. That is fine for many and disqualifying for some companies. Know which you are.
Using these in OpenClaw
OpenClaw runs each of these models two ways: through the vendor's own provider plugin, on that vendor's API key or coding plan, or through one OpenAI-compatible provider block in ~/.openclaw/openclaw.json that serves all of them on one key. Either way you set a model ref (provider/model) under agents.defaults.model.primary and can list fallbacks behind it. Cheapest first, at list prices:
| Model | OpenClaw plugin and key | Model ref | Typical day |
|---|---|---|---|
| DeepSeek V4.1 Flash | @openclaw/deepseek-provider, DEEPSEEK_API_KEY | deepseek/deepseek-flash | $0.22 off-peak, $0.43 peak |
| GLM 5.3 Flash | @openclaw/zai-provider, ZAI_API_KEY | zai/glm-5.3-flash | $0.41 |
| MiniMax M3 | Bundled minimax plugin, MINIMAX_API_KEY or Coding Plan sign-in | minimax/MiniMax-M3 | $0.84 |
| GLM 5.3 | @openclaw/zai-provider, API key or GLM Coding Plan | zai/glm-5.3 | $3.70 |
| Kimi K3 | @openclaw/moonshot-provider, MOONSHOT_API_KEY; Kimi Code uses @openclaw/kimi-provider | moonshot/kimi-k3 (kimi/k3 on Kimi Code) | $6.25 |
Plugins and refs from OpenClaw's provider docs, checked October 14, 2026. Typical day = the median coding-agent day in our Agent Burn Index at each vendor's per-token list price; OpenClaw is not broken out there, so it stands in, and a plan subscription changes the math.
One key for all of them in OpenClaw
Through Standard Compute, the same models sit in one standardcompute provider block ("baseUrl": "https://api.stdcmpt.com/v1", "api": "openai-responses") with refs such as standardcompute/deepseek/deepseek-v4.1-flash or standardcompute/z-ai/glm-5.3; "mode": "merge" keeps OpenClaw's built-in providers. Back up the file and run openclaw config validate before restarting the Gateway. Full configs: GLM 5.3 in OpenClaw and DeepSeek in OpenClaw.
Cheapest is not the whole answer. OpenClaw's security docs recommend the strongest model for bots with tools, because cheaper tiers are more open to prompt injection, and GLM 5.3 scores highest of these on Terminal-Bench 4.0. A common setup runs a stronger model as primary and a cheap one for heartbeats. Best LLM for OpenClaw ranks the options, and cutting OpenClaw's API costs covers the other levers.
The catch
You are not buying a discounted frontier model; you are buying a very good open-weight model with tight operational limits from vendors that reprice aggressively. Priced-in honestly, that is still an excellent deal, as the volume layer of a two-layer setup rather than your only model.
Where Standard Compute fits
Standard Compute's router does the two-layer setup automatically: efficient open models (GLM, DeepSeek, Qwen, Kimi and MiniMax among them) handle routine requests and frontier models such as Claude Fable 5.1 and GPT-6 Astra take the hard ones, on providers we vet for price, uptime and speed with automatic fallback, under one fixed monthly budget from $19/mo. Opt-in Europe or US Compliance rules keep a key, team or company on providers in that region, and we don't train on your prompts or store them.
FAQ
Is the GLM Coding Plan worth it?
At $18/mo Lite: yes for volume routine work, if you accept the credit windows (2,000 credits per 5 hours and 10,000 a week on Lite) and the supported-tools-only terms. The community-standard setup pairs it with a frontier option for hard tasks rather than using it alone.
Are these plans safe to use for company code?
That is a policy question, not a quality one: the plans work, but your code transits infrastructure under Chinese jurisdiction. Many companies are fine with it, many compliance teams are not. Ask yours before wiring it into work repos.
Which is best: GLM, Kimi or MiniMax?
They leapfrog each other with every model release; differences are smaller than their shared catches (windows, concurrency, peak slowdowns). Pick on current model quality for your language/stack, and assume you may switch; the community does, often.
What is the cheapest model for OpenClaw: Kimi, MiniMax, GLM or DeepSeek?
DeepSeek V4.1 Flash: about $0.22 for a typical coding-agent day off-peak and $0.43 at peak, at DeepSeek's list prices. GLM 5.3 Flash is close at $0.41, then MiniMax M3 at $0.84, GLM 5.3 at $3.70 and Kimi K3 at $6.25. Cheapest is not best for an agent with tools: GLM 5.3 scores highest of the five on Terminal-Bench 4.0, and OpenClaw's security docs recommend the strongest model for tool-enabled bots.
How do I configure Kimi, MiniMax, GLM or DeepSeek in OpenClaw?
Install the vendor's plugin and run onboarding: openclaw plugins install @openclaw/deepseek-provider, @openclaw/zai-provider or @openclaw/moonshot-provider (MiniMax's plugin is bundled), then openclaw onboard with the matching auth choice, such as deepseek-api-key, zai-coding-global for the GLM Coding Plan, moonshot-api-key or minimax-global-oauth. Switch later with openclaw models set and a ref such as zai/glm-5.3. To use all of them on one key, add a single OpenAI-compatible provider block to ~/.openclaw/openclaw.json instead.
Can I use a GLM, Kimi or MiniMax coding plan in OpenClaw?
OpenClaw has onboarding for each: zai-coding-global for the GLM Coding Plan, the Kimi Code provider (kimi/ refs) for Kimi's plan, and an OAuth sign-in for MiniMax's plan. Read each plan's terms first: GLM's plan is for supported coding tools and Kimi Code for interactive use, so an always-on OpenClaw bot may not fit them.