News / Model Watch · GLM promotion
GLM-5.3-Flash Runs Free Overnight Through Sep 20: Zhipu's Build Events Push the Off-Peak-Free Era Deeper
Zhipu is giving away its flash tier's quiet hours. Per Sina Tech (Sep 5), the official ZCode activity makes GLM-5.3-Flash free and unlimited every night from 23:00 to 09:00 (Beijing) through September 20, adds a 300-million-token weekend drop (Saturdays 09:00, first-come), and auto-credits new users with 100M tokens. After DeepSeek's weekend off-peak billing in August, this is the second major vendor to price its idle hours at zero — with a catch that matters: the free tokens live only inside ZCode, not the API.
Global Build (Sep 3–20): nightly 23:00–09:00, GLM-5.3-Flash unlimited and free inside ZCode; Coding Plan subscribers get the 10 hours quota-free — reportedly the first Zhipu drop where paid users went first. Weekend Build: Saturdays 09:00, 300M tokens per person, expires Sunday 23:00, first-come. New users: 100M tokens on signup; Max tier: reported 8.2B weekly flash cap. The model is the former「牛来」/ Ox Alpha — 320B MoE (18B active), tri-modal, AA index 57 ≈ Claude Opus 4.8 at ~1/40 the API price. API rate cards unchanged — see the GLM API guide for production pricing.
What the events give, item by item
- Global Build (started Sep 3, runs 18 days to Sep 20): every night 23:00–09:00 Beijing time, GLM-5.3-Flash is free and unlimited inside ZCode — a 10-hour daily window.
- Weekend Build: Saturdays at 09:00, 300M GLM-5.3-Flash tokens per person, first-come-first-served, expiring Sunday 23:00. Earlier 100M-token drops sold out in under an hour; the second round's extra 50,000 seats went just as fast.
- Who gets priority: GLM Coding Plan subscribers get the nightly 10 hours with zero quota burn — per the report, the first time paid users were prioritized over new users. Max-tier subscribers carry a reported weekly 8.2B-token flash cap ("basically impossible to exhaust," per Zhipu's community lead quoted by the report).
- How to claim: ZCode desktop client v3.10+ (macOS/Windows/Linux), activity card in the lower-left corner, official entry zcode.z.ai. Free tokens are usable only inside ZCode.
- Confidence note: Tier B — activity details per Sina Tech (Sep 5, via AI信息Gap) citing Zhipu's official activity card; treat exact caps and windows as of Sep 5.
Why it matters: the off-peak-free pattern
- Nights at ¥0 is the new off-peak. DeepSeek priced weekends at valley rates in August; Zhipu now prices nights at zero in September. Both are the same message: flash-tier serving has idle capacity worth burning for adoption.
- The subsidy targets the harness, not the API. Free tokens locked to ZCode (not API keys) extend the strategy around GLM-5.3 — model + coding harness + subscription — where retention comes from workflow lock-in rather than token price. For API builders, list rates didn't move.
- Flash-tier pressure keeps building. The model being given away scores level with Claude Opus 4.8 on Artificial Analysis' index at ~1/40 the price — the same flash tier where Gemini 3.8 Flash just entered the price war. Free nights are another leg on that ladder.
- Real arbitrage, if your loop fits the window: batch jobs (overnight evals, doc processing, agent sweeps) scheduled 23:00–09:00 inside ZCode run at zero token cost through Sep 20. Production API traffic — the thing our relay comparison tracks — is untouched.
- Caveats: single secondary source for exact numbers; event windows end Sep 20; ZCode-only means no OpenAI-compatible endpoint for the free quota.
Primary sources
FAQ (2026)
What's free and when?
Sep 3–20: nightly 23:00–09:00 Beijing time, GLM-5.3-Flash unlimited in ZCode — 10 hours a day for 18 days.
The weekend drop?
Saturdays 09:00, 300M tokens per person, first-come, expires Sunday 23:00. Earlier 100M rounds emptied in under an hour.
Is the API free too?
No — free usage is ZCode-client-only. API rate cards stand; relays and z.ai endpoints bill normally.
Who's eligible?
Everyone for nightly free; new users +100M tokens; Coding Plan subscribers get quota-free nights (paid users first); Max tier 8.2B/week cap.
What's GLM-5.3-Flash?
320B MoE (18B active), tri-modal, the former「牛来」/ Ox Alpha. AA index 57 ≈ Claude Opus 4.8 at ~1/40 the API price.
Why care as a builder?
Off-peak free is now a pattern (DeepSeek weekends → GLM nights): idle flash capacity is being burned for adoption. Batch jobs in the window are real arbitrage.