Three generations · independent explainer

Subscription plan generations · not model generations

The plan everyone remembers as three dollars a month is gone. Here is exactly what replaced it, and what that costs you.

Z.ai’s has been rebuilt three times in twelve months. Two of those three generations can no longer be bought at any price, and the two that survive as “legacy” are treated very differently from one another — a distinction almost nobody makes, and the one that decides what you should do next.

All figures verified 2026-08-14 · 51 dimensions · 45 sources

  • qualified“Legacy V1 was the most generous generation”
  • confirmed“Legacy V1 is being phased out”
  • corrected“The legacy plans are all going away”
  • qualified“There are three generations”

Each of those verdicts is argued out in the section below, with the sources that decide it.

I hold a legacy plan I’m new to all this

Entry price is the cheapest tier at the time. Select a generation to highlight it in every table on this page.

Orientation

If you have never heard of any of this

Four short answers, and then you can read the rest of the page like everyone else.

What a GLM Coding Plan actually is

Coding assistants like Claude Code, Cline and Roo Code talk to a language model over an API. Normally you pay that model’s vendor per request. A GLM Coding Plan is a flat monthly subscription that lets those same tools talk to Z.ai’s GLM models instead, at a fraction of the cost.

You change one setting — the address your coding tool points at — and your usage comes out of the subscription rather than off a card.

The catch newcomers miss This is not an API plan. The subscription may only be used inside about nineteen named coding tools. Calling it from your own application, bot or website is explicitly prohibited in the subscription terms and can get your benefits restricted.

Why three generations is a real problem, not trivia

Each generation counts your usage in a different unit. The first two counted — one thing you ask is one prompt, no matter how much work it sets off. The current one counts , derived from .

That means the numbers you find in blog posts and comparison sites are frequently describing a plan that no longer exists, in a unit the current plan does not use. Several pricing sites we checked on 2026-08-14 still present the retired prompt figures as current.

Every row in the tables below says which generation it belongs to and how confident we are in it. That is the whole point of this page.

What the money means for a working day

The cheapest current tier is $18 a month, or $12.60 if you pay for a year. Z.ai’s own estimate is that it buys 43 to 87 million a week — the range depends on how much of your work lands in , when everything costs twice as much.

For most developers outside Asia that range’s top end is realistic, because peak is 06:00–10:00 UTC on weekdays and weekends are cheap all day.

Push a real workload through it →

How today’s offer compares to the high-water mark

In late 2025, $3 a month bought about 120 queries every five hours with no at all. Today’s $18 tier has a weekly ceiling and meters differently, so the two cannot be compared directly — but on price alone the entry point is six times what it was.

Z.ai used to advertise the plan as “3× usage, 1/7 cost”. It makes no equivalent claim now.

See it row by row →

Every term on this page, defined
cache hit rate

The share of the text sent to the model that it has already seen and stored, and so charges less for.

Z.ai states 90.9% as the average for coding workloads, and uses that figure for its own token estimates. Cached input costs about a quarter of fresh input on GLM-5.3.

concurrency

How many requests you can have in flight at the same time — which decides whether you can run two projects, or a swarm of sub-agents, at once.

Z.ai has never published a number for this. It says only that the limits follow Max > Pro > Lite and are adjusted dynamically according to available capacity.

context window

How much text the model can hold in mind at once — your code, the conversation and everything it has read.

The current plan offers up to a million tokens on every tier, but you have to opt in by adding a suffix to the model name.

credit

A budget measured in text. Everything the model reads and writes is converted into credits by a published formula, with the model’s output costing far more than its input.

Credits = (input tokens × multiplier + cached input × multiplier + output × multiplier) ÷ 10,000. On GLM-5.3 the multipliers are 6.9, 1.7 and 24, so one output token costs about fourteen times what one cached input token costs.

GLM Coding Plan

A monthly subscription from Z.ai that lets you use its GLM models inside coding tools such as Claude Code, Cline or Roo Code, instead of paying per request.

You point your coding tool at a Z.ai endpoint instead of the model vendor’s own, and your usage comes out of the subscription rather than off a card. It is not a general API plan: using it from your own application is a terms violation.

grandfathered

Keeping the terms you originally signed up on after those terms stop being offered to new customers.

Here it is conditional in both directions: the first generation’s privilege required auto-renew to be switched on before a deadline, and Z.ai’s contract separately reserves the right to stop renewing anyone at all.

peak hours

A window when using the newest models costs more of your allowance. On this plan it is weekday afternoons in Singapore.

Monday to Friday, 14:00–18:00 Singapore time (UTC+8) — twenty hours a week. That is 06:00–10:00 UTC, so many developers outside Asia are almost never in it. Weekends are off-peak all day.

pro-rated

Splitting a charge by how much of the period you actually used.

On this plan the value that comes back from a pro-rated upgrade arrives as account credit, never as money returned to your card.

prompt quota

A budget measured in questions. One thing you ask counts as one prompt, no matter how much work the assistant does answering it.

Z.ai’s own definition: “One prompt refers to one query. Each prompt is estimated to invoke the model 15–20 times.” So a single prompt could cover a long chain of file reads, edits and test runs. The first two generations counted this way; the current one does not.

rate window

The stretch of time your allowance is measured over. Use it all inside the window and you wait for the window to refresh.

Every generation has used a five-hour window. On the current plan the refresh is rolling: each credit becomes available again five hours after you spent it, rather than everything resetting at one moment.

token

A chunk of text, roughly three-quarters of a word. Models read and write in tokens, and the current plan bills by them.

A medium source file might be two or three thousand tokens. A coding assistant re-sends much of the conversation on every turn, which is why cached input dominates the total.

weekly cap

A second, longer ceiling on top of the short one. You can be well inside your five-hour allowance and still be stopped for the rest of the week.

Introduced on 12 February 2026. The first generation had no weekly cap at all, which is the single biggest reason it is remembered as the generous one.

The claim under test

What people say about these plans, and whether it holds up

This page started from a widely-shared framing: that the first generation was the generous one, and that it is being phased out. We treated that as something to check rather than something to repeat.

“Legacy V1 was the most generous generation”Mostly true, with two real catches

True of V1 as it settled, and by a wide margin: about 120 user queries every five hours, no weekly ceiling at all, and new models arriving at no extra cost. One researcher measured that allowance at 40 million tokens every five hours — more, on the entry tier, than the current entry tier gets. Two catches, though. The famous $3 was a first-month price; the recurring price was $6, and Z.ai’s marketing went on quoting the $3 for months. And the famous 120 was not the launch figure either: the page captured on launch morning advertised 15 to 50, and was revised upward within hours. The generous V1 people remember is real, but it is a specific window, and it was never quite as cheap as the number everyone repeats.

Sources:

“Legacy V1 is being phased out”True, and further along than “being”

Z.ai cancelled auto-renewal on every eligible V1 plan on 30 April 2026 and said legacy plans “will no longer remain as a supported subscription option going forward”. There is no renewal path. Every V1 subscription still alive today is running out its last paid term. It is not being phased out; it has been.

Sources:

“The legacy plans are all going away”No — the two legacy cohorts are treated very differently

V2 is not on the same path. Z.ai’s July 2026 announcement explicitly lets V2 holders keep auto-renew switched on and upgrade tier or billing term. It offers no such right to V1. One legacy generation is terminal; the other can be renewed indefinitely, at least as of today. Anyone lumping “legacy” together is missing the single sharpest fact in the story.

Sources:

“There are three generations”Three is the right count, but only two are numbered

Z.ai names “Legacy Plan V1” and “Legacy Plan V2” in your account. The current one has no version number at all — the company just calls it the credits-based plan. “V3” is a community label. And V1 changed materially inside its own lifetime, so if you counted by terms rather than by Z.ai’s labels you could argue for four.

Sources:

The spine

All three generations, side by side

51 dimensions across 3 generations. Sort it, filter it, hide what you do not care about, highlight one generation, or put two side by side and show only the change. Every cell carries a confidence badge and a numbered source.

Wording
Highlight
Every dimension that decides something, across all three generations. Expert wording. Exact units, exact terms. Each figure carries its confidence and a numbered source you can open. Verified 2026-08-14.
DimensionWhat is being compared Legacy Plan V1Sold new: 1 September 2025 – 12 February 2026 Legacy Plan V2Sold new: 12 February 2026 – 30 July 2026 The credits planOn sale since 30 July 2026
What Z.ai calls itThe label that appears in your account under “My plan”.V1“Legacy Plan V1” OfficialV2“Legacy Plan V2” OfficialCredits“The new credits-based plan” — no version number“V3” is a community shorthand. Z.ai numbers only the two retired generations. Official
Can you buy it today?New subscriptions, not renewals.V1No — withdrawn from sale, and auto-renew was force-cancelled OfficialV2No to new buyers — but existing holders may still renew OfficialCreditsYes Official
Sold to new customers duringWhen you could walk up and buy it.V11 Sep 2025 → 12 Feb 2026Announced at 14:27 UTC on 1 September 2025 in a single post on X, with no launch blog post. The July 2025 date sometimes quoted belongs to the GLM-4.5 model, not to this subscription. OfficialV212 Feb 2026 → 30 July 2026 OfficialCredits30 July 2026 → today Official
What ended the previous generationThe event that closed the door behind each generation.V1— (the first generation) OfficialV2A cutoff at 12 Feb 2026 (UTC+8): subscribe and enable auto-renew before it and you kept V1 terms for the life of the subscription Archived officialCreditsThe 30 July 2026 announcement stopped selling the prompt plans and opened the credits plan Official
Team versionTeam seats have always been metered on their own scheme.V1Individual plans onlyThe July 2026 notice tags both legacy cohorts “individual plans only”. OfficialV2Individual plans only OfficialCreditsStandard Seat $88/seat/mo, Premium Seat $188/seat/mo, minimum 2 seatsTeam seats meter in raw tokens, not credits, and are the only place overage billing exists. Official
Lite — monthly list priceThe cheapest way in.V1$3 for the first billing cycle, $6 recurringThe famous $3 was a first-month offer, not the standing price. Z.ai’s own marketing kept saying “starting at just 3 USD per month” for months afterwards, which is how the number lodged itself in everyone’s memory. The recurring price was double that. Archived officialV2$10 / month, then $18 from around 19 April 2026Two distinct price points inside one generation. The first-purchase discount disappeared at the same time, so a new buyer’s real cost went from $3 to $10 overnight. Archived officialCredits$18 / month Official
Pro — monthly list priceThe tier most working developers land on.V1$15 for the first billing cycle, $30 recurringSame structure as Lite: the advertised figure was a first-month price and the recurring price was double it. Archived officialV2$30 / month, then $72 from around 19 April 2026Pro was the one tier the February change left alone — Z.ai’s announcement named only Lite and Max. It then went up by 140% in April. Archived officialCredits$80 / monthAlso reported: $72 / month. Z.ai’s April 2026 migration notice printed $72 for Pro, and a third-party tracker still repeats it. That was the previous generation’s price. We read $80 as the struck-through list price on the live pricing page on 2026-08-14. Official
Max — monthly list priceThe heaviest individual tier.V1Not offered at launch. Priced at twice Pro — so $30 first cycle, $60 recurring.Z.ai stated the two-times-Pro ratio officially; the dollar figures come from its archived checkout data. The tier appears in the documentation by 1 October 2025, a month after launch. Archived officialV2$80 / month, then $160 from around 19 April 2026A user watching it happen put it at $80 two weeks before 20 April and $160 on the day, which matches the archived checkout data to within a few days. Archived officialCredits$168 / monthAlso reported: $160 / month. $160 was the previous generation’s Max price and is still repeated by third-party trackers. Official
Discount for paying up frontWhat you save by committing to a quarter or a year.V1No standing term discount. Annual was simply twelve months at list, with seasonal promotions on top.A Christmas 2025 promotion took 20% off annual on top of the first-purchase discount, which is how an annual Lite plan came to cost $28.80. Archived officialV2Quarterly −10%; annual −30%, cut to −20% at the April price riseStanding discounts appear for the first time in this generation. The annual discount was then reduced in the same move that raised the prices, so the annual cost rose by more than the headline monthly figure suggests. Archived officialCreditsQuarterly −20%, yearly −30%. Lite works out at $14.40 or $12.60 a month.Team seats get only 10%, not 30%. No page anywhere shows the total you actually pay up front — only the effective monthly rate. Official
Priced in another currency?The mainland-China sibling sells the same thing in yuan.V1Not established UnverifiedV2Not established UnverifiedCreditsYes: ¥118, ¥538 and ¥1,078 a month on bigmodel.cn, with identical allowancesAt the yuan prices the tiers are not a straight currency conversion of the dollar ones. The Chinese pricing page also defaults to quarterly rather than yearly, so a careless comparison mixes two different billing terms. Official
Cheapest real monthly costLite, on the longest commitment available.V1$2.40 / month — $28.80 for a year, in the December 2025 promotionTwenty percent off an annual Lite plan that already carried a first-purchase discount. This is the cheapest the GLM Coding Plan has ever been. Archived officialV2$7 / month early on ($84 a year), rising to $14.40 ($172.80 a year) Archived officialCredits$12.60 / month on yearly billingThe effective monthly rate shown on the pricing page at the yearly −30% setting. Official
Discounts on offerThings that reduce the sticker price.V1Introductory pricing was itself the promotion Archived officialV2A 50% migration discount for V1 holders, from 30 Apr 2026 until three months after their legacy term endedIt floats: it is always half of whatever the current discounted price is, not a frozen figure, and it could be used repeatedly inside the window. OfficialCredits10% off a first order via a referral link; the old 50% migration discount still works hereThe referrer only gets paid after three successful invites, which is easy to miss. Official
What gets countedThe unit your allowance is denominated in.V1Prompts. One prompt = one user query, estimated to invoke the model 15–20 times. Archived officialV2Prompts, same definition OfficialCreditsCredits = (input×mult + cached input×mult + output×mult) ÷ 10,000This is the single biggest change between generations. Under prompts, a runaway agent loop inside one query cost you one prompt. Under credits it costs whatever it burns. A captured response from Z.ai’s undocumented quota endpoint suggests the older plans were typed internally as token limits all along, so the prompt counts may always have been a presentation layer over the same thing. Official
Five-hour allowance, LiteHow much you can use in one sitting on the cheapest tier.V1~120 prompts — measured at about 40M tokensThe token figure is a community measurement, cross-checked against Z.ai’s own quota endpoint while that endpoint still returned real numbers. It is the only measurement anywhere that lets the generations be compared in one unit. Archived officialV2~80 prompts — about 27M tokens on the same measurementA third less than before. The token equivalent is our arithmetic: the cut was exactly two thirds across all three tiers. Archived officialCredits2,000 credits ≈ 8.6M tokens at peak, 17.2M off-peakDerived by scaling Z.ai’s own weekly token estimate for Lite down to the five-hour allowance. Notice what this implies: measured against the first generation’s 40 million, the current entry tier’s sitting allowance is smaller. Derived
Five-hour allowance, ProSame question, one tier up.V1~600 prompts Archived officialV2~400 prompts OfficialCredits12,000 credits Official
Five-hour allowance, MaxSame question, top tier.V1~2,400 prompts Archived officialV2~1,600 prompts OfficialCredits28,000 credits Official
Weekly ceilingThe cap that decides whether a heavy week is possible at all.V1NoneZ.ai’s own shorthand for the first generation was “legacy plans (without weekly usage limits)”. OfficialV2Lite ~400, Pro ~2,000, Max ~8,000 prompts per week — exactly five times the five-hour figureThe weekly ceiling landed days after the grandfathering cutoff, not on it. One dated community record says it arrived at four times the five-hour allowance on 14 February and was loosened to five times on the 16th — so the very first version was 20% tighter than the figures that ended up in the documentation. OfficialCreditsLite 10,000, Pro 60,000, Max 140,000 credits per week Official
Most you could use in a week, LiteThe ceiling if you never stopped.V1~4,000 prompts ≈ 1,344M tokens — no weekly stop, only the five-hour oneOur arithmetic: 40 million tokens per five-hour window, 33.6 windows in a week. No real person sustains that; it is the shape of the ceiling, not a forecast. DerivedV2~400 prompts ≈ 133M tokens — the weekly cap binds firstAnd that is before the model multiplier: on the newest models a token counted two or three times, so the real ceiling was closer to 45–65M tokens of flagship use. DerivedCredits10,000 credits ≈ 43–87M tokensZ.ai’s own estimate, assuming all use is on its newest model at a 90.9% cache-hit rate. The range is peak versus off-peak. Read against the row above, the current entry tier lands close to the second generation’s effective ceiling and far below the first generation’s. Official
Z.ai’s own value claimWhat the company said you were getting.V1“About 3× the usage quota of the Claude Pro plan”, at “only ~1% of standard API pricing” Archived officialV2“Approximately 15–30× the monthly subscription fee” in API-equivalent value OfficialCreditsNo equivalent multiplier is claimedAbsence of evidence: we found no comparable multiple on any current page. The credit table and token estimates replaced the marketing claim. Derived
Search and browse tool allowanceThe extra tools that sit alongside the model.V1Not documented in the snapshots we recovered UnverifiedV2Lite 100, Pro 1,000, Max 4,000 web search + reader calls per monthThis detail survives only on the archived April 2026 page. Archived officialCredits1.2 credits per call, from the same pool as everything else Official
Short windowThe one that bites during a long session.V15 hours Archived officialV25 hours OfficialCredits5 hours, refreshed rolling: each credit returns 5 hours after it was spentA rolling refresh, not a clock-aligned bucket. You are never waiting for a fixed reset time. Official
Weekly resetWhen the weekly counter goes back to zero.V1Not applicableThere was no weekly counter to reset. OfficialV27-day cycle OfficialCredits7 days from the moment you subscribed — not Monday Official
Peak-hours penaltyUsage costs more in a narrow weekday window, Singapore time.V1Not documented for this generationPeak coefficients first appear in the second generation’s documentation. Absence of evidence, not evidence of absence. UnverifiedV2Newest models cost 3× at peak and 2× off-peakAlso reported: 3× peak / 1× off-peak. This is how Z.ai restated the same generation’s terms in July 2026, after it had ended. The pages subscribers actually read at the time said 2× off-peak, with 1× offered as a limited-time benefit that was extended repeatedly — “valid through the end of April”, then “through the end of June”. The two official texts do not agree, and the difference is a factor of two on most people’s usage. ContestedCreditsOff-peak costs 50% of the standard credit rateSame window, reframed: the second generation anchored on off-peak and penalised peak; this one anchors on peak and rewards off-peak. The ratio also narrowed from 3:1 to 2:1. Official
When peak actually isThis matters more than it sounds.V1Not documented UnverifiedV214:00–18:00 Singapore time, every dayAlso reported: Monday to Friday only. The weekdays-only qualifier appears in the July 2026 restatement, not in anything a subscriber read at the time. The contemporary pages say simply “peak hours are 14:00–18:00 (UTC+8)”, with no day restriction — so weekend afternoons appear to have been charged at peak rates during this generation. ContestedCreditsMon–Fri 14:00–18:00 SGT; all weekend is off-peakTwenty hours a week are peak. That window is 06:00–10:00 UTC — so a developer in Europe or the Americas may almost never be in it. Official
How many things at onceParallel sessions and sub-agents.V1Published as numbers on launch day: 10 concurrent requests on Lite, 30 on ProThose figures were on the pricing page on 1 September 2025 and were quietly removed within days. No later generation has published a number. Archived officialV2Numbers withdrawn. In practice one user observed a limit of one request in flight.Also reported: Officially: no number at all, only “Max > Pro > Lite”. Z.ai replaced the published figures with an ordering and a dynamic-adjustment clause. A Pro subscriber then documented reaching only 4% of his five-hour quota before hitting concurrency errors. One report is not a specification — but there is no specification to check it against, which is the point. ContestedCreditsNo published number. Guidance only: Lite one project, Pro 1–2, Max 2+, with Max > Pro > Lite and higher limits off-peakZ.ai says the limits are adjusted dynamically according to available capacity, so there is no figure to hold them to. This may matter more than any allowance on this page: if throughput binds first, the ceilings never come into it. Official
SpeedHow fast tokens actually arrive, which can bind before any quota does.V1Not documented UnverifiedV2Users reported congestion during peak and slower throughput than Claude CommunityCreditsPro advertises “faster generation speeds”, Max “dedicated resources during peak times” — no figuresImplies the cheapest tier may be speed-limited, but Z.ai publishes no throughput figures for any tier. Official
Models you can callWhat the subscription actually gets you access to.V1GLM-4.5 era modelsThe 2025 snapshot advertises GLM-4.5 in coding tools. Model line-ups shifted repeatedly during this plan’s life. Archived officialV2GLM-5.x and GLM-4.x models of the day Archived officialCreditsGLM-5.3, GLM-5-Turbo, GLM-4.7 — and nothing elseRequests for older models are silently routed to GLM-5.3. Calling anything outside the three produces a balance error. Official
Do cheaper tiers get worse models?A question most people assume the answer to.V1No. Every tier got every model, and new models arrived at no extra cost.GLM-4.5 became 4.6 became 4.7 with no price change. That is a real and under-remembered part of why this generation is remembered fondly. Archived officialV2Yes, for a while. Lite was excluded from GLM-5 at launch and calling it returned an error; resolved by May 2026.The only period in which model access genuinely depended on tier. Users who had bought an annual Lite plan expecting the newest model were, in their own words, not pleased. Archived officialCreditsNo. All three tiers reach all three models.Higher tiers get earlier access to new models and more capacity, not different models. Official
Context windowHow much of your codebase the model can hold at once.V1Not documented per tier UnverifiedV2Not documented per tier UnverifiedCreditsUp to 1,000,000 tokens, opt-in with a [1m] model suffix. Same for every tier.Context is a property of the model, not the tier. One Z.ai config page still shows an older model name for this setting. Official
Search, browsing and visionThe tools bundled alongside the model.V1Not documented in the snapshots we recovered UnverifiedV2Web search and reader, on a monthly count Archived officialCreditsVision understanding, web search, web reader and Zread — on every tierThe pricing page’s Pro card implies these are a Pro feature; two documentation pages say all plans get them. We follow the documentation. Official
Where you may use itThe restriction newcomers trip over most.V1Inside supported coding tools Archived officialV2Inside supported coding tools OfficialCreditsOnly inside ~19 named tools. General API access from your own code is a terms violation.This is not an API plan. Calling it from your own app, bot or website is explicitly prohibited and may get benefits restricted. Official
Behaviour at the limitBlocked, throttled, queued, or billed.V1Blocked until the five-hour window refreshed Archived officialV2Blocked until the relevant window refreshed OfficialCreditsHard block. No overage, no fallback to pay-per-call.Z.ai’s FAQ only mentions waiting for the five-hour cycle. If it is the weekly ceiling you hit, waiting five hours does nothing — you wait for the seven-day reset. Official
Can you pay to keep going?An escape hatch, or the absence of one.V1Not documented UnverifiedV2Not documented UnverifiedCreditsIndividual: no. Team seats: yes, opt-in overage at API list price less 10%.The 10% overage discount is described as a limited-time offer with no stated end date. Official
Does it eat your account balance?A common worry.V1Not documented UnverifiedV2Not documented UnverifiedCreditsNo. Plan calls only ever use plan quota. Official
Auto-renewalWhether the plan renews itself.V1Force-cancelled for all eligible holders on 30 April 2026Applied to holders who had auto-renew enabled on that date. Z.ai never says what happened to holders who had already switched it off. OfficialV2Still available: holders may keep auto-renew on and keep renewing OfficialCreditsOn by default at the end of each cycleThe legal terms describe renewal as something you opt into at checkout; the plan usage policy describes it as automatic. Two official pages, two framings. Official
Is your price locked?The assumption most grandfathered subscribers are quietly making.V1Terms held for the life of the subscription — but the subscription was ended Archived officialV2Terms hold to the end of each billing cycle OfficialCreditsNo. Renewal is charged at the price in force on the charge date.The contract says this outright, and adds that continuing to use the service after a price change counts as accepting it. There is no price-lock guarantee anywhere. Official
Can Z.ai simply stop renewing you?The clause that made the first generation’s ending lawful.V1Yes — and it did OfficialV2Yes, the same clause applies OfficialCreditsYes: Z.ai may “unilaterally discontinue providing automatic renewal services” as its operational strategy requires, with noticeThis is the single most important sentence for anyone counting on a grandfathered price. Official
Cancellation noticeHow long before your billing date you must act.V1Same general terms OfficialV2Same general terms OfficialCreditsAt least 3 days before the billing date — on two Z.ai pagesAlso reported: 24 hours before the end of the current term. A genuine two-against-two split inside Z.ai’s own documentation. The usage policy and the pricing-page FAQ say at least three days; the docs FAQ and the legal subscription terms say twenty-four hours. Cancel three days out and you cannot be caught by either reading. Contested
RefundsGetting money back.V1None OfficialV2None OfficialCreditsNone, even for unused quotaThree official pages agree. The only carve-out is where law requires one. Official
Upgrading mid-termWhat happens to the time you already paid for.V1Cannot switch to a current plan before expiry OfficialV2May switch early only by upgrading a tier; the old plan ends immediately and its unused value goes toward the new priceA one-way door. Once a V2 holder upgrades early, V2 is gone. OfficialCreditsCross-tier upgrade takes effect immediately with pro-rated value returned as account balance. Same-tier changes stack instead: monthly to annual gives you 13 months.The thirteen-month outcome surprises people. Pro-rated value comes back as account credit, never as cash. Official
DowngradingMoving to a cheaper tier.V1Not available OfficialV2Queued to the end of the current term OfficialCreditsTakes effect after the current cycle ends. No proration. Official
Sharing the subscriptionOne plan, how many people.V1Same general terms OfficialV2Same general terms OfficialCreditsLicensed to one named person. Sharing, renting, reselling or proxying is prohibited; enforcement runs rate limiting → freeze → ban after three violations.The restriction is on multiple users, not multiple devices. One person on several machines is nowhere prohibited. Official
Is your code used to train models?Worth knowing before you point it at a private repository.V1Same general terms OfficialV2Same general terms OfficialCreditsFor individual users, yes: Z.ai reserves the right to process your prompts and the model’s outputs to develop and improve its models, and you consent to it. API and enterprise customers get the opposite default.We found no opt-out documented for individuals — only an invitation to contact them. The Coding Plan is sold to individuals, so this is the relevant default. Official
Liability capWhat you can recover if it goes wrong.V1Same general terms OfficialV2Same general terms OfficialCreditsCapped at your spend in the most recent calendar month Official
Can a current holder keep renewing?The question that decides everything else.V1No. Auto-renew was cancelled on 30 April 2026 and the renewal right is absent from the July 2026 notice.The July 2026 notice lists a renewal right for V2 and conspicuously does not list one for V1, which matches April’s force-cancellation. Z.ai never states the negative outright. OfficialV2Yes. “If you currently have a legacy plan V2, you can enable auto-renew or upgrade your plan.” OfficialCreditsYes — it is the plan on sale Official
What forces you offThe specific events that end a grandfathered plan.V1Expiry of the current paid term. Nothing else is needed — renewal was already removed. OfficialV2Letting it lapse, or upgrading a tier early — which ends the plan immediatelyZ.ai never states what happens if you lapse. Since the plan is “no longer sold to new users”, the strong reading is that a lapse is irreversible. That inference is ours. OfficialCreditsNothing — it is the current plan Official
If you lapse, can you get it back?The irreversibility question.V1No OfficialV2Almost certainly not — but Z.ai never says so directlyInference from “previous plans are no longer sold to new users” plus a renewal right described as available only “while your plan is active”. No Z.ai sentence states the consequence of lapsing. DerivedCreditsNot applicable — you can just subscribe again Official
Can you pause it?For a holiday, a job change, a quiet quarter.V1No pause feature is documented DerivedV2No pause feature is documented — and cancelling is a one-way exitWe searched the usage policy, FAQ and subscription terms. Cancel is the only lever, and it ends renewal. DerivedCreditsNo pause feature is documented Derived
What you were given in exchangeMigration support offered when each generation closed.V1Two complimentary months of the matching current tier, granted automatically on 30 April 2026 and added after the existing term; plus 50% off until three months after the legacy term ended, reusable within that window.Tier-matched and stacked after the existing term rather than replacing it. The 50% is a floating half of the current price, not a frozen figure. OfficialV2Nothing beyond keeping the plan — which, since it stays renewable, is arguably the better deal OfficialCreditsNot applicable Official
Closest thing on sale todayWhere a holder of a closed generation would land.V1Lite at $18/month for the price, or Pro at $80/month to get near the old headroom. Neither restores an uncapped week. DerivedV2The same tier of the credits plan. Z.ai says the generations “differ only in how usage is calculated”.Z.ai’s claim that only the calculation changed is a claim, not a measurement. Nobody has published a conversion between prompts and credits. OfficialCreditsYou are already on it Official
What your old money buys nowThe same monthly spend, priced at today’s list.V1$3/month buys no plan today. The cheapest entry is $18, or $12.60 on yearly billing. DerivedV2$18/month still buys Lite — the same nominal price, a different unit of account. DerivedCredits$18/month buys Lite; $80 Pro; $168 Max Official
How to read the confidence badges Official means it is on a Z.ai page that is live today. Archived official means it was on a Z.ai page that has since changed, and we read it from a dated archive. Community means users reported it and Z.ai did not. Derived means we did arithmetic on cited figures. Contested means sources disagree and we show both. Unverified means we looked and could not find out, which we would rather say than guess.

Twelve months

Scrub through what happened

Drag the handle, or use the arrow keys. 14 dated moments from launch to today — three price rises, one quota cut, one grandfathering deadline and two sunsets.

Read the whole timeline as a list
  1. 1 Sep 2025The GLM Coding Plan launches, on X and nowhere elseZ.ai announces it at 14:27 UTC with a single post: one seventh the price of the Claude Code plans, three times the prompts. Two tiers — Lite at $3 for your first month, Pro at $15 — reverting to $6 and $30 afterwards. Access to GLM-4.5 and GLM-4.5-Air.There is no launch blog post; the announcement is the post. The July 2025 date sometimes quoted belongs to the GLM-4.5 model, not to this subscription.
  2. Same dayThe allowance is revised eightfold upward within hoursThe page captured at 11:20 UTC advertises about 15–50 queries per five hours on Lite and 80–240 on Pro, alongside hard concurrency figures — 10 requests at once on Lite, 30 on Pro. By 17:31 the same day it reads about 120 and 600.Both captures are of Z.ai’s own page, hours apart. The 120 figure is the one that stuck and the one that held until February 2026 — but the generous generation’s opening position lasted about six hours. The concurrency numbers were dropped within days and no later generation ever published one again.
  3. Oct 2025Newer models arrive at the same priceGLM-4.6, and later GLM-4.7, become available to existing subscribers without a price change. This is a genuine and often-forgotten part of why the first generation is remembered fondly.
  4. Dec 2025A Max tier, a Christmas deal, and a sloganA third tier is documented at about 2,400 queries per five hours. The site banner reads “3× usage, 1/7 cost”, and the docs claim the plan delivers roughly three times a Claude allowance at about one percent of standard API pricing.Community reports from the same period describe a Black Friday offer of a full year of Pro for around $120 — about $10 a month.
  5. 11 Feb 2026Z.ai announces price rises — and only price risesA post on X: first-purchase discounts removed, Lite and Max going up, existing subscribers keeping their pricing. A companion post adds, plainly, “To be upfront: compute is very tight.” Pro’s list price is untouched, exactly as announced.What the announcement does not mention is the thing that mattered more. The 33% cut to the five-hour allowance and the brand-new weekly ceiling were never announced anywhere — they arrived as an edit to a documentation page.
  6. 12 Feb 2026The grandfathering cutoffSubscribe and switch auto-renew on before this date, Singapore time, and your original quota holds for the life of the subscription with no weekly limit. Miss it and you are on the new terms.This is the boundary between the two legacy generations, in Z.ai’s own words. Note what the privilege was conditional on: auto-renew already being enabled. On the same day, Z.ai’s quota endpoint stopped returning token counts and began returning only percentages — which is why nobody has been able to measure the plan precisely since.
  7. 13–16 Feb 2026The allowance is cut and a weekly ceiling appearsLite drops from about 120 queries per five hours to 80, Pro from 600 to 400, Max from 2,400 to 1,600. Then a weekly cap lands on top: 400, 2,000 and 8,000 queries — exactly five times the five-hour figure.One dated community record says the weekly cap arrived on the 14th at four times the five-hour allowance and was loosened to five times on the 16th. Either way it landed after the grandfathering deadline had already passed, so nobody weighing that deadline knew what they were choosing between.
  8. 7 Mar 2026An apology, and fifteen days of creditAfter weeks of reliability complaints, Zhipu emails subscribers on the China side compensation equivalent to fifteen days of their subscription fee, with no expiry.Reposted and translated by users rather than published, and it is not clear whether international subscribers received the same. It corroborates the congestion complaints appearing on forums in the same period.
  9. Early Apr 2026First price rise landsThe documentation now reads “starting at just 10 USD per month, with Pro plans from 30 USD per month”. With the first-month offer gone, a new buyer’s real entry cost has gone from $3 to $10.
  10. 12 Apr 2026Second price rise: everything roughly doubles againLite reaches $18, Pro $72, Max $160 — and the annual discount is cut from 30% to 20% in the same move, so an annual buyer is hit twice. No announcement was made at all.Z.ai’s own checkout data shows the same product identifiers carrying the new prices. A user watching it happen wrote that Max “was $80 two weeks ago, and now it’s $160”; another that a three-month Lite plan had gone from $27 to $48.
  11. 21 Apr 2026The Legacy Plan Migration NoticeZ.ai announces that legacy plans — the ones without weekly limits — will be phased out and will not remain a supported option. It acknowledges, in its own words, that this lands on top of a recent price increase.
  12. 30 Apr 2026V1's auto-renewal is switched off for everyoneExisting paid periods run to their end, but no V1 subscription can renew. In compensation, eligible holders get two complimentary months of the matching current tier and 50% off until three months after their legacy term ends.The 50% is a floating half of whatever the current discounted price is, and it can be used repeatedly inside the window — so a V1 holder could stack a long term at half price before the door closed.
  13. 30 Jul 2026Everything switches to creditsThe unit of account changes from queries to tokens, converted into credits by a published formula. Both prompt-based generations stop being sold. Z.ai names them Legacy Plan V1 and Legacy Plan V2 for the first time.The asymmetry is set here: V2 holders may keep auto-renew on and upgrade; V1 holders may do neither and must wait for expiry.
  14. 14 Aug 2026Where things stand as we checkedThe credits plan is on sale at $18, $80 and $168 a month. V2 remains renewable with no announced end date. Every remaining V1 subscription is counting down to its final billing date. GLM-5.3 became the flagship model on the plan today, with the credit multipliers unchanged.Everything on this page was verified on this date. Z.ai has changed plan terms three times in twelve months, so treat anything here as perishable — and note that most third-party comparisons you will find are describing a generation that no longer exists, in a unit the current plan does not use.

Burn it down

Push your week through all three generations at once

Describe how you actually work. We run that week against each generation’s ceilings and show you which ones stop you. This is the only place on the page where we bridge between and , and that bridge is an assumption you can move.

Start from

6
7 One thing you ask — however many tool calls it sets off — counted as one query.
5
20% Peak is 06:00–10:00 UTC on weekdays. Outside Asia-Pacific office hours, a low number here is realistic.
100,000 The bridge between the two metering systems. Z.ai has never published one. Move it and watch the current plan’s verdict change.
Legacy Plan V1———
Legacy Plan V2———
The credits plan———

Your week, in numbers

Queries a week
—
Queries in the heaviest five hours
—
Tokens a week
—
Credits a week on the current plan
—

Which tier is being tested

Comparing like with like: the same tier name across all three generations. The first generation never published a Max price, so that cell stays empty rather than being invented.

Which model you lean on

This matters more than it looks. The first generation charged the same for everything; the second charged up to three times for its newest models; the current one charges less for the older one and halves everything off-peak.

Every assumption behind those bars, and why
Tokens per user query — 250,000 by default, adjustableTwo independent community measurements bracket this. One researcher measured the first generation’s Lite tier at 40 million tokens per five hours against a published ceiling of 120 queries, giving about 333,000 tokens per query. A separate user reported staying around 20 to 25 million tokens across 100-plus queries, giving 200,000 to 250,000. Z.ai has never published this bridge. It is the biggest lever here, which is why you can move it.
Credits per million tokens — 232.6 at peak, 116.3 off-peakBack-solved from Z.ai’s own published pair — 10,000 weekly credits on Lite, estimated at 43 to 87 million tokens a week at a 90.9% cache-hit rate. The same constant then predicts Pro at 258 million against a stated 263, and Max at 602 against a stated 613, so it holds across all three tiers rather than being fitted to one.
First-generation ceiling in tokens — 40 million per five hours on LiteMeasured by a community researcher in January 2026 and cross-checked against Z.ai’s quota endpoint, which returned real token counts until 12 February 2026 — the day the second generation launched, after which it returned only percentages. Pro and Max are scaled from Lite by the published prompt ratios, which is our arithmetic.
Second-generation ceilings — Two thirds of the first, weekly at five times the windowThe quota cut was exactly two thirds across all three tiers, and the weekly ceiling was set at five times the five-hour ceiling — 400 against 80, 2,000 against 400, 8,000 against 1,600. Applying both ratios to the measured token figure gives the numbers used here.
Model multipliers — 1× / 3× / varies by generationThe first generation had no multipliers of any kind. The second charged three times at peak and twice off-peak for its newest models. The current one charges half rate off-peak and about a third less for the cheaper model. This is why the same week can fit one generation and overrun another at identical raw token counts.
Peak share of your work — Your sliderPeak is Monday to Friday, 14:00 to 18:00 Singapore time — twenty hours a week, and 06:00 to 10:00 UTC. A developer in Europe or the Americas may almost never be in it, so a low number here is realistic for most of the world. Note that the second generation’s documentation did not say weekdays only; that qualifier appears in the later restatement.
What this deliberately ignores — ConcurrencyOne Pro subscriber documented reaching only 4% of his five-hour quota before hitting concurrency errors, and observed a practical limit of one request in flight. If that is representative, throughput binds long before any ceiling here does — and fanning work out across sub-agents does not burn quota faster, it fails instead. No generation ever published a concurrency number, so we cannot model it.

Cross-generation comparison requires a bridge that Z.ai has never published. We have made ours visible and adjustable rather than burying it. If the bridge is wrong, the current generation’s bars move and the two legacy ones do not.

Calibrate yourself

How generous was it really?

Five figures. Guess each one before you see it. Most people are wrong in the same direction, and the gap between what you guessed and what is true is the actual lesson.

Question 1 of 5 0 close

 

I’d rather just read the answers
  • ?The plan launched in September 2025 advertising $3 a month. What did it actually cost from month two onward?Answer: $6
  • ?How many user queries could a Lite subscriber send every five hours by late 2025?Answer: 120
  • ?After February 2026, how many queries per five hours did the same Lite tier get?Answer: 80
  • ?What share of the week counts as expensive “peak” time on the current plan?Answer: 12%
  • ?Between early and late April 2026, the Max tier went from $80 a month to what?Answer: $160

For newcomers

Which tier, if any, fits you

Five questions about how you actually work. The answer is sometimes “none of them”, and we will say so.

How much of your working day has a coding assistant in the loop?

Be honest about the busy weeks rather than the quiet ones — the ceilings bite on the busy ones.

For anyone holding a legacy plan

What you have, what you are about to lose, and what to do about it

Four questions. The answer is a reasoned verdict with the trade-offs stated, not a recommendation to buy anything. If holding is right, it will say hold. If leaving is right, it will say that too.

Which plan are you on right now?

Your account shows this under “My plan”. If it says Legacy Plan V1 or Legacy Plan V2, that is your answer.

How much you have uncovered

The picture so far

Nothing here is required. Everything on this page is readable without earning a single one of these — they are a record of what you have actually looked at.

  • OrientedRead what the plan actually is
  • Table readerUsed a table control
  • Difference spotterCompared two generations directly
  • Source checkerOpened a citation
  • Jargon busterLooked up a definition
  • Time travellerScrubbed the whole timeline
  • Burn testerRan a workload past a ceiling
  • CalibratedFinished the generosity guesses
  • MatchedCompleted the plan matcher
  • DecidedGot a personal assessment

Take your result with you

Built from what you did on this page. Nothing is sent anywhere — this page makes no network requests at all.

Explore the page and your summary will build here.

What to do next

  • Holding V1: find your renewal date. It is the date your plan stops existing. Decide before it, not after — there is no path back.
  • Holding V2: you are the cohort that can still renew. Do not upgrade a tier casually; an early upgrade ends the legacy plan immediately and permanently.
  • On the credits plan: check what share of your work is inside 06:00–10:00 UTC. Moving work out of it halves what it costs you.
  • Considering it: confirm your coding tool is on the supported list first, and read the training clause if the code is not yours.

Then check the figures yourself. Every one of them has a number beside it that opens the source.

Show your working

Every source, in full

45 of them. Live Z.ai pages, dated archive snapshots of Z.ai pages that have since changed, and dated community reports where no official record survives.

  1. GLM Coding Plan pricing pagehttps://z.ai/subscribe OfficialZ.airead 2026-08-14The live pricing page, read in a real browser in all three billing-term states on 2026-08-14. It is client-rendered, so a plain fetch returns no prices at all. It also opens on the Yearly tab, which is why so many write-ups quote $12.60 as “the monthly price” — it is the effective monthly rate at the yearly −30% setting.
    “$12.6/month $18/month … $56/month $80/month … $117.6/month $168/month”
  2. The same plan on bigmodel.cn, in Chinesehttps://docs.bigmodel.cn/cn/coding-plan/overview OfficialZhipu / bigmodel.cnread 2026-08-14Z.ai’s mainland-China sibling runs the identical credits system: the same three models, the same five-hour and weekly credit allowances, the same hard stop when they run out. Prices differ — ¥118, ¥538 and ¥1,078 a month — and its pricing page defaults to quarterly rather than yearly. A few details, such as an auto-renew ceiling and an uplift on legacy team plans, appear only in the Chinese notices.
    “套餐类型 5 小时积分 每周积分 / Lite 套餐 2,000 10,000 / Pro 套餐 12,000 60,000 / Max 套餐 28,000 140,000”
  3. GLM Coding Plan — plan overview and credit tablehttps://docs.z.ai/devpack/overview OfficialZ.ai docsread 2026-08-14The primary specification page for the current credits-based plan: credit allowances, the credit formula, per-model multipliers, peak hours and the token estimates.
    “Each plan is subject to both a 5-hour usage limit and a weekly usage limit.”
  4. GLM Coding Plan FAQhttps://docs.z.ai/devpack/faq OfficialZ.ai docsread 2026-08-14Covers what happens at the limit, upgrade and downgrade mechanics, cancellation and refunds.
    “Once the quota is used up, you’ll need to wait until the next 5-hour cycle for it to refresh.”
  5. GLM Coding Plan usage policyhttps://docs.z.ai/devpack/usage-policy OfficialZ.ai docsread 2026-08-14Concurrency guidance, the renewal and cancellation rules, account-sharing prohibition and the enforcement ladder.
    “Rate (concurrency) limits are tied to your plan tier … the general principle being Max > Pro > Lite.”
  6. Plan Update Announcement — the move to creditshttps://docs.z.ai/devpack/notice/usage-revision OfficialZ.ai docspublished 2026-07-30The launch document for the current generation, and the only page where Z.ai names its own grandfathered cohorts “Legacy Plan V1” and “Legacy Plan V2”. It also tabulates the superseded prompt quotas, which is why those figures are quotable at all.
    “The new credits-based plan is now available. Previous plans are no longer sold to new users.”
  7. Legacy Plan Migration Notice — the V1 sunsethttps://docs.z.ai/devpack/transition OfficialZ.ai docspublished 2026-04-21The document that ended the first generation: auto-renew cancelled on 2026-04-30, plus the two complimentary months and the 50% migration discount offered in compensation.
    “Legacy plans (without weekly usage limits) will be phased out and will no longer remain as a supported subscription option going forward.”
  8. Switching to the latest model / 1M contexthttps://docs.z.ai/devpack/latest-model OfficialZ.ai docsread 2026-08-14Model availability across tiers and how the one-million-token context is switched on.
    “To enable GLM 1M context, add the [1m] suffix to the model name (e.g., glm-5.3[1m])”
  9. Supported tools and API endpointshttps://docs.z.ai/devpack/tool/others OfficialZ.ai docsread 2026-08-14The list of coding agents the subscription may legally be used inside, and the base URLs each needs.
    “Incorrect endpoint configuration will result in inability to use GLM Coding Plan subscription quota.”
  10. Subscriptions, Fees, and Payment (legal terms)https://docs.z.ai/legal-agreement/subscription-terms OfficialZ.ai legalread 2026-08-14The binding contract. Contains the clause that pre-authorises exactly what happened to the first generation, and the clause that defeats any assumption of a locked-in price.
    “the price for automatic renewal shall be the price … actually applied by the system on the date the charge is made, rather than the original price at the time of your initial subscription.”
  11. Terms of Usehttps://docs.z.ai/legal-agreement/terms-of-use OfficialZ.ai legalread 2026-08-14Defines User Content as your prompts and the model’s outputs, and sets a different default for individuals than for API and enterprise customers.
    “For individual users, we reserve the right to process any User Content to improve our existing Services and/or to develop new products and services”
  12. Team Plan benefitshttps://docs.z.ai/devpack/teamplan OfficialZ.ai docsread 2026-08-14The separate team track, which meters differently again and is the only place overage billing exists.
    “A minimum of 2 seats is required, with no upper limit on the number of seats”
  13. Credit campaign and referral ruleshttps://docs.z.ai/devpack/credit-campaign-rules OfficialZ.ai docslast updated 2026-03-15Referral discounts and their conditions. Note the page’s own date predates the July replatform.
    “the final amount payable after applying all discounts and Credits must be at least 0.50 USD”
  14. Plan overview as it stood on 2025-09-09 (archived)https://web.archive.org/web/20250909015313/https://docs.z.ai/devpack/overview Archived officialInternet Archive / Z.ai docssnapshot 2025-09-09The clearest surviving statement of first-generation terms, on Z.ai’s own page, before it was rewritten. This is where the $3 entry price and the 120-prompt allowance come from.
    “Lite Plan: Up to ~120 prompts every 5 hours — about 3× the usage quota of the Claude Pro plan.”
  15. Plan overview as it stood on 2025-12-23 (archived)https://web.archive.org/web/20251223043504/https://docs.z.ai/devpack/overview Archived officialInternet Archive / Z.ai docssnapshot 2025-12-23The first snapshot in which a Max tier appears alongside Lite and Pro in the first generation.
    “Max Plan: Up to ~2400 prompts every 5 hours — about 3× the usage quota of the Claude Max (20x) plan.”
  16. Plan overview as it stood on 2026-03-10 (archived)https://web.archive.org/web/20260310070146/https://docs.z.ai/devpack/overview Archived officialInternet Archive / Z.ai docssnapshot 2026-03-10The grandfathering promise itself, in Z.ai’s own words. It fixes the generational boundary at 2026-02-12 and makes the privilege conditional on auto-renew already being switched on.
    “For users who subscribed and enabled auto-renewal before February 12 (UTC+8), the original quota will remain in effect throughout the subscription validity period, and no weekly usage limits will apply.”
  17. Plan overview as it stood on 2026-02-13 (archived)https://web.archive.org/web/20260213210131/https://docs.z.ai/devpack/overview Archived officialInternet Archive / Z.ai docssnapshot 2026-02-13Captured the day after the cutoff. Shows the second generation’s lower five-hour ceiling already live.
    “Lite Plan: Up to ~80 prompts every 5 hours”
  18. Plan overview as it stood on 2026-03-13 (archived)https://web.archive.org/web/20260313094957/https://docs.z.ai/devpack/overview Archived officialInternet Archive / Z.ai docssnapshot 2026-03-13The $3 entry price was still printed here a month after the quota change, which is why snapshot dates are upper bounds on when a change actually took effect rather than the change date itself.
    “Starting at just 3 USD per month, with Pro plans from 15 USD per month”
  19. Plan overview as it stood on 2026-04-04 (archived)https://web.archive.org/web/20260404102848/https://docs.z.ai/devpack/overview Archived officialInternet Archive / Z.ai docssnapshot 2026-04-04Catches the first of the two 2026 price rises in flight.
    “Starting at just 10 USD per month, with Pro plans from 30 USD per month”
  20. Plan overview as it stood on 2026-04-29 (archived)https://web.archive.org/web/20260429132114/https://docs.z.ai/devpack/overview Archived officialInternet Archive / Z.ai docssnapshot 2026-04-29Second-generation detail that no longer exists anywhere live: monthly counts for the search tools, and the off-peak coefficient as it was worded before the current page.
    “Lite Plan: Include a total of 100 web searches and web readers per month”
  21. Z.ai’s own pricing endpoint, captured 2025-12-23https://web.archive.org/web/20251223123822id_/https://api.z.ai/api/biz/pay/batch-preview Archived officialInternet Archive / api.z.aisnapshot 2025-12-23The checkout API returned every plan’s list price, discounted price and campaign name as JSON, and the archive caught it. This is where the first generation’s real list prices come from — and it shows that the famous $3 was a first-purchase discount off a $6 list, not the standing price.
    “Lite $6→$3, Pro $30→$15, Max $60→$30 monthly, all via “First Purchase Discount””
  22. Z.ai’s pricing endpoint, captured 2026-03-01, 03-27 and 04-05https://web.archive.org/web/20260327195211id_/https://api.z.ai/api/biz/pay/batch-preview Archived officialInternet Archive / api.z.aisnapshots 2026-03-01 to 2026-04-05Three captures agreeing on the second generation’s first pricing: Lite $10, Pro $30, Max $80 a month, with a standing 10% quarterly and 30% annual discount. Two of the three were captured from a session that was not subscribed, so these are the prices a new buyer saw.
    “"campaignName":"-30% per Year"”
  23. Z.ai’s pricing endpoint, captured 2026-04-19https://web.archive.org/web/20260419181845id_/https://api.z.ai/api/biz/pay/batch-preview Archived officialInternet Archive / api.z.aisnapshot 2026-04-19The same product identifiers now carry the higher prices — Z.ai raised prices on existing items rather than creating new ones — and the annual discount has been cut from 30% to 20% in the same move.
    “"monthlyOriginalAmount":18.00,"monthlyPayAmount":14.40 … "campaignName":"-20% per Year"”
  24. Z.ai announces the GLM Coding Plan on Xhttps://x.com/Zai_org/status/1962522757536887205 Official@Zai_org2025-09-01The launch, at 14:27 UTC on 1 September 2025. There was no launch blog post — this post is the announcement. The July 2025 date sometimes given belongs to the GLM-4.5 model, not to the subscription.
    “Announcing GLM Coding Plan for Claude Code! … 1/7th the price of original Claude Code plans - 3x more prompts.”
  25. Cline’s write-up of the plan at launchhttps://cline.bot/blog CommunityCline2025-09A third-party integrator spelling out what the pricing page’s “1st Month Offer” badge meant: the famous $3 was a first-billing-cycle price and the recurring price was $6. We could not deep-link the individual post, so this points at the blog index — but the same fact is independently visible in Z.ai’s archived checkout data, which shows $6 and $30 as the list prices with a “First Purchase Discount” applied on top.
    “these prices are for the first billing cycle after which the price will increase to $6/$30”
  26. Z.ai’s price-change announcement on Xhttps://x.com/Zai_org/status/2021656635668901985 Official@Zai_org2026-02-11The only announcement Z.ai made for the February change, and it covered price alone. A companion post said plainly: “To be upfront: compute is very tight.” The quota cut and the brand-new weekly ceiling were never announced anywhere — they arrived as silent edits to the documentation.
    “Existing subscribers keep their current pricing.”
  27. Plan overview as it stood on 2026-02-06 (archived)https://web.archive.org/web/20260206230642/https://docs.z.ai/devpack/overview Archived officialInternet Archive / Z.ai docssnapshot 2026-02-06The last archived page showing first-generation terms. The word “weekly” appears nowhere on it. Together with the 13 February capture, it brackets the change to a single week.
    “Lite Plan: Up to ~120 prompts every 5 hours”
  28. Plan overview as it stood on 2026-02-18 (archived)https://web.archive.org/web/20260218174544/https://docs.z.ai/devpack/overview Archived officialInternet Archive / Z.ai docssnapshot 2026-02-18The second generation’s specification in its settled form: the two-column quota table, the grandfathering sentence, and the peak multiplier as it was originally written — three times at peak and twice off-peak, not the one-times that later documentation claims.
    “Its usage will be deducted at 3 × during peak hours and 2 × during off-peak hours.”
  29. Plan overview as it stood on 2026-05-15 (archived)https://web.archive.org/web/20260515135237id_/https://docs.z.ai/devpack/overview.md Archived officialInternet Archive / Z.ai docssnapshot 2026-05-15Late second-generation state. Shows the one-times off-peak rate as a temporary benefit with an expiry date, repeatedly extended — which is why the later claim that off-peak was simply one-times does not match what subscribers were told at the time.
    “As a limited-time benefit, GLM-5.1 and GLM-5-Turbo will only consume 1× quota during off-peak hours, valid through the end of June.”
  30. The usage policy, first archived 2026-03-16https://web.archive.org/web/20260316185324id_/https://docs.z.ai/devpack/usage-policy Archived officialInternet Archive / Z.ai docssnapshot 2026-03-16This page did not exist during the first generation. Its arrival mid-second-generation is itself a finding: account sharing rules, non-coding-use restrictions and a three-strikes ban policy all appear for the first time here.
    “it may be subject to risk control measures, including high-intensity throttling, account suspension, or permanent ban”
  31. Pricing-page FAQ content, captured 2026-04-23https://web.archive.org/web/20260423115543id_/https://api.z.ai/api/biz/operation/query?ids=1136 Archived officialInternet Archive / api.z.aisnapshot 2026-04-23The pricing page’s FAQ text, captured through the API that feeds it. Documents the one period when model access really was gated by tier, and the list of supported tools at the time.
    “The Lite plan currently does not include GLM-5 quota… If you call GLM-5 under the plan endpoints, an error will be returned.”
  32. Hacker News: Max tier doubled in about two weekshttps://news.ycombinator.com/item?id=47838634 CommunityHacker News (user Kerrick)2026-04-20A dated user observation that brackets the second price rise more tightly than the archive can.
    “It was $80 two weeks ago, and now it’s $160.”
  33. Hacker News: quarterly price moved $27 to $48, and a weekly cap hit on day onehttps://news.ycombinator.com/item?id=47855184 CommunityHacker News (user UncleOxidant)2026-04-21Two useful things in one comment: an independent price observation, and the only concrete report we found of a full week’s prompt allowance being burned inside a single day.
    “I must’ve somehow hit their weekly limit on day one of the week.”
  34. Hacker News: 7 million tokens consumed 2% of a weekly quotahttps://news.ycombinator.com/item?id=47708846 CommunityHacker News (user recursivegirth)2026-04-09Implies roughly 350 million tokens a week on that tier, which brackets Z.ai’s own current estimate for Pro. A rare case of a user measurement and an official figure agreeing.
    “7 million tokens has only gone through 2% of my weekly usage”
  35. Hacker News: $180 for a year, bought in Decemberhttps://news.ycombinator.com/item?id=48182676 CommunityHacker News (user rescbr)2026-05-18A first-generation annual price seen in the wild. The tier is not stated, so treat the mapping to Pro as inference rather than fact.
    “Last December I paid $180 for an year of Z.ai’s coding plan.”
  36. Hacker News: off-peak is where the value sitshttps://news.ycombinator.com/item?id=48108363 CommunityHacker News (user brokegrammer)2026-05-12Confirms the off-peak discount was live and widely understood during the second generation.
    “During off-peak hours, usage for GLM-5.1 uses only 1x of your quota”
  37. Hacker News: tighter limits after the price risehttps://news.ycombinator.com/item?id=48097115 CommunityHacker News (user aspectrr)2026-05-11Subjective, and the published quota numbers did not change in that window, so this most likely describes throughput and concurrency rather than quota.
    “I like the GLM coding plan before they raised their prices, now their rate limits are more strict”
  38. Hacker News: good value, slower throughputhttps://news.ycombinator.com/item?id=48631479 CommunityHacker News (user nijave)2026-06-22Relevant because wall-clock speed, not quota, may be what actually limits a heavy day.
    “z.ai seems a bit on the slower side for raw model tok/sec throughput.”
  39. Hacker News: plan abuse degraded service in early Aprilhttps://news.ycombinator.com/item?id=47641008 CommunityHacker News (user mariopt)2026-04-04The most commonly offered explanation for why caps and prices moved. Z.ai never gave a reason, and the token figure the poster cites is an unrelated marketplace total, so treat the causal story as plausible rather than established.
    “Last week Z.ai coding plan was unusable due to a lot of people abusing the coding plan”
  40. jia.je — a dated knowledge base of Chinese coding planshttps://jia.je/kb/en/software/coding_plan.html CommunityJiegec (jia.je)read 2026-08-14An unusually rigorous community record: dated entries, stated methodology, explicit “tested” versus “speculated” labels, and figures for the current plan that match Z.ai’s own documentation exactly. It is the only source anywhere that measured what a prompt was worth in tokens, cross-checked against Z.ai’s own quota endpoint before that endpoint stopped returning numbers.
    “speculated that GLM Coding Plan’s Lite plan usage limit is that the sum of all requests' input + output tokens does not exceed 40M per 5 hours (meaning each prompt corresponds to 40M/120=333K tokens)”
  41. GitHub issue: a Pro subscriber reaching 4% of quota before being rate-limitedhttps://github.com/anomalyco/opencode/issues/8618 CommunityGitHub2026-01-15The most consequential community finding on this page. Z.ai endorses sub-agents and says concurrency follows Max > Pro > Lite, but publishes no numbers; this subscriber observed a practical limit of one request in flight. If that is representative, fanning work out across sub-agents does not burn quota faster — it fails instead.
    “I consistently hit 'Too much concurrency' errors and can only utilize approximately 4% of my 5-hour quota before being rate-limited.”
  42. GitHub issue: a captured payload from Z.ai’s undocumented quota endpointhttps://github.com/robinebers/openusage/issues/1104 CommunityGitHub2026-08-13A real captured response from an endpoint Z.ai does not document. Two things fall out of it: Z.ai does internally version the current plan as V3, and the previous generation was typed as a token limit — which suggests the published prompt counts were always a presentation layer over a token budget.
    “New Z.ai coding plans (subscription entries report "version": "V3" …) return quota entries with "type": "CREDIT_LIMIT" instead of "TOKENS_LIMIT"”
  43. Reddit: a heavy Lite session measured in tokenshttps://www.reddit.com/r/ZaiGLM/comments/1r587il/glm_47_lite_plan_getting_frequent_5_hour_quota/ Communityr/ZaiGLMaround February 2026The closest thing to a citable heavy-coding hour anywhere. Its implied 200,000 to 250,000 tokens per prompt independently brackets the 333,000 measured elsewhere. Reddit shows only relative dates, so the date is approximate.
    “even with 100+ prompts, I usually stayed around ~20–25M tokens and never hit the 5-hour rate limit”
  44. codingplan.org tracker entry for GLMhttps://codingplan.org/en/plans/glm Communitycodingplan.orgread 2026-08-14A third-party tracker whose weekly-credit figures match the official docs exactly, which is why it is quoted here as corroboration. It still names an older model as current, so it lags by weeks.
    “quarterly 20% and yearly 30% discounts”
  45. Arithmetic performed for this page#methodology DerivedThis explainer2026-08-14A figure produced by combining cited numbers rather than one Z.ai published. The working is set out in the methodology note, and the inputs are cited beside it.

The small print

How this page was built, and where it will go wrong

This is an independent explainer It is not affiliated with, endorsed by, or reviewed by Z.ai. Nothing here is advice about what to buy. Check anything that matters against the source links before you spend money on it.

What we did

We read every live Z.ai page describing the plan — the pricing page, the plan overview, the FAQ, the usage policy, the team plan page, the two migration announcements and the legal subscription terms. For the retired generations, whose terms no longer exist anywhere live, we worked from dated Internet Archive snapshots of Z.ai’s own documentation, which is server-rendered and therefore readable in the archive. The pricing page is not: from late September 2025 onward the archive captured only an empty shell, so archived prices come from the documentation rather than from the pricing page.

Where no record survives at all, we used dated community reports and labelled them as such. Where two sources disagree, we show both readings rather than picking one quietly.

What we are least sure about

  • The exact day of the first price rise, which we could only bracket to a two-week window in March or early April 2026. The second rise is pinned to 12 April by a dated changelog and two corroborating user reports.
  • The first generation’s Max tier price in dollars. Z.ai stated the ratio — twice the Pro price — officially, but the dollar figures come from archived checkout data rather than from a published price list.
  • Whether a lapsed legacy plan can ever be repurchased. Z.ai never addresses it. We infer no from “previous plans are no longer sold to new users”, and we label that inference.
  • Concurrency on any generation. It has never been published as a number.
  • The bridge between prompts and credits in the simulator. That is our assumption, not Z.ai’s, and it is adjustable for exactly that reason.

Where sources contradict each other

  • Cancellation notice. Two Z.ai pages say at least three days; two say twenty-four hours. A clean two-against-two split. Use three days and you cannot be wrong.
  • What a prompt was worth. The English announcement estimates each legacy prompt at 15–20 model calls; the Chinese version of the same notice says 15–30.
  • Auto-renewal. The plan usage policy describes it as automatic; the legal terms describe it as something you opt into at checkout.
  • The off-peak coefficient on the second generation. An archived April 2026 page said 2× off-peak with a temporary 1× offer; the live page says a flat 1×.
  • Whether V1 holders can renew. One summary line in the July 2026 announcement says holders of plans discontinued on 30 July can still renew — but the V1-specific section grants no renewal right, and V1 was discontinued in April, not July. Read with April’s force-cancellation, the summary line appears to be about V2. We flag it rather than resolve it.

What could change after 2026-08-14

Z.ai has changed plan terms three times in twelve months and its contract reserves the right to change prices, to amend the agreement on publication, and to “unilaterally discontinue providing automatic renewal services” as its operational strategy requires. Renewals are charged at the price in force on the day the card is charged, not the price you originally signed up at. Treat every figure on this page as true on the verification date and perishable after it.

View more demos Get up to 40% off GLM-5.3