Blog / Service updates

GLM Coding Plan Quota Explained: Credits, Peak Hours and Reset Rules

2026-10-19cdnping editors4 min read

GLM Coding Plan (from Zhipu, the Chinese AI lab behind the GLM models) now meters usage in "credits," with two layers of limits: a 5-hour cap and a weekly cap. On weekdays from 14:00 to 18:00 Beijing time (UTC+8), model calls are deducted at full price; at all other times they are deducted at 50%. Everything below is based on the official Zhipu Coding Plan documentation as we verified it on October 17, 2026. The page does not date its rule changes, so we can't say when they took effect. Treat the official page as the source of truth.

#How many credits does each tier get?

Both limits apply at the same time. Hitting either one blocks further requests.

#How many credits does one call cost?

The official formula:

Credits consumed = (input tokens × Input coefficient + cache-hit tokens × Cached Input coefficient + output tokens × Output coefficient) ÷ 10,000

MCP usage is "number of calls × Output coefficient," so each call costs 1.2 credits. Two things stand out: output tokens carry the highest coefficient, and cache-hit input is much cheaper than uncached input. In a long conversation, a higher cache hit rate means the same credits last longer. Our prompt caching cost-saving guide explains how to check cache hits (it covers Claude, and the principle is similar, but Zhipu's documentation governs the actual behavior).

#An example (assumed numbers)

Suppose one GLM-5.3 request carries 100,000 input tokens, of which 95,000 hit the cache, and produces 3,000 output tokens:

• (5,000 × 6.9 + 95,000 × 1.7 + 3,000 × 24) ÷ 10,000 = 26.8 credits
• Off-peak, at 50%, that is 13.4 credits

Against Lite's 2,000 credits per 5 hours, that works out to roughly 74 such requests at peak and about 149 off-peak. Real agent tasks vary a lot in tokens per turn, so this only demonstrates the arithmetic.

#When are the peak hours?

Monday to Friday, 14:00–18:00 (UTC+8). The documentation says that outside peak hours, model calls are deducted at 50% of the base credits. The discount applies only to model calls, and the documentation lists no time-of-day discount for MCP.

Other providers have similar off-peak rules, for example DeepSeek API peak and off-peak pricing. It's worth comparing how the billing logic differs.

#When does the quota reset?

• 5-hour credits: a rolling refresh. Credits come back 5 hours after the request that consumed them.
• Weekly credits: the cycle starts at the time of purchase and repeats every 7 days.
• Upgrading your plan: credits refresh immediately to the new tier's amounts.
• Renewing: does not reset credits, and the existing refresh cycle stays the same.
• Running out: you wait for the next 5-hour recovery. The plan does not draw on other resource packs or your account balance.

So "my quota got reset out of nowhere" or "it ran out after only a short session" is usually down to the rolling window and the day your weekly cycle started. To find your own cycle, check the order date; the cycle repeats every 7 days from there. For how other products handle limits, see how Codex usage limits work and Claude Pro/Max usage limit resets.

#How do older models and third-party tools fit in?

Automatic model switches listed in the official docs:

• GLM-5.2 and GLM-5.1 are switched to GLM-5.3
• GLM-5-Turbo and GLM-4.7 are switched to GLM-5.3-Flash

For background, Zhipu's Chinese changelog lists GLM-5.3 as released on August 19, 2026 and GLM-5.3-Flash on August 26. The English Z.AI changelog gives August 18 for GLM-5.3; the one-day gap may be a time zone difference.

Plan quota only works in the officially supported tools. The documentation lists ZCode, AutoClaw, Claude Code, Codex, OpenCode, WorkBuddy and others. Calling the API outside those tools does not draw on the plan quota. OpenClaw can be used but is treated as secondary scheduling, so requests may queue or be throttled under heavy load.

#How to spend fewer credits: a checklist

1. Schedule big refactors, batch file edits and long-context analysis before 14:00, after 18:00 or on weekends.
2. Use GLM-5.3-Flash for simple tasks (completion, formatting, classification). Its coefficients are roughly a third of GLM-5.3's.
3. Keep sessions continuous so the cache hits. Switching context often makes input count at full price.
4. Don't let agents call web-search-type MCP tools in a loop. Each call is 1.2 credits, and loops add up.
5. When you're close to the 5-hour cap, wrap up the current task instead of starting a new large one.

#Rate limited: is it you or the service?

First separate two cases. Quota exhaustion means your own credits ran out; a service problem means other people are seeing errors too. Check cdnping's Zhipu API status (probed from the edge every 5 minutes), then the Zhipu page on the "is it down" tracker, which collects user reports such as rate limiting and "quota got reset." If status looks normal and nobody else is reporting problems, you have most likely used up your credits. For how to read provider status pages, see how to read AI status pages. You can also follow documentation and pricing changes on the changes tracker.

#Sources

• Zhipu GLM Coding Plan overview (verified 2026-10-17)
• Zhipu Open Platform changelog
• Z.AI Release Notes