GLM Coding Plan 2026 Pricing and Limits Explained: How to Choose a Plan

Z.ai's GLM Coding Plan has become the default choice for many developers running coding agents day to day — starting at $18/month, and by Z.ai's own calculations, its weekly quota covers 48 million to 97 million tokens of the flagship GLM-5.3 model. If you bought the same usage at Z.ai's own pay-as-you-go prices, that would come to roughly $70 to $140 per month.

But the problem is that the plans described in many reviews no longer exist. Within 2026 alone, the plan went from "counted by prompt, no weekly cap" to "counted by prompt, with 5-hour and weekly caps," and then on July 30 it switched entirely to a token-metered credits system. Prices have also moved up: the Pro tier went from $72/month in April to $80 today.

This article breaks down the mechanics the pricing page doesn't make clear: how credits are actually deducted, how many dollars each tier's quota is worth, and where the break-even point sits compared with pay-as-you-go and third-party hosting.

1. What Is the GLM Coding Plan

The GLM Coding Plan is a monthly subscription from Z.ai (the international brand of Zhipu AI) for calling GLM models within AI coding tools using a fixed quota, rather than paying per token. You create an API Key, point supported agents like Claude Code and Cline at Z.ai's coding endpoint, and each call consumes fixed quota rather than drawing on a credit card balance.

The three tiers include the same models — the flagship GLM-5.3 and the faster GLM-5.3-Flash — plus four MCP tools: image understanding, web search, web reading, and repository reading. The tiers differ only in usage, and a single quota is shared across all connected tools.

This plan is not an API plan: Z.ai's subscription terms explicitly prohibit users from calling the models "from your own applications, bots, websites, SaaS products, or other systems." Quota cannot be shared with colleagues, does not fall back to account balance when exhausted, and is non-refundable.

Why Fixed-Rate Plans Exist

Coding agents are a special kind of customer for model providers. Every step of an agent loop re-sends the prior conversation (system prompt, tool definitions, files already read), so a single task can push millions of tokens, the vast majority of which are identical to the previous call. Z.ai charges $1.40 per million tokens for fresh input on GLM-5.3, but only $0.26 per million for cached input — under pay-as-you-go, heavy users face enormous bills that are reread-dominated and hard to predict.

So the plan suits three types of people: developers who use agents heavily every day; users who want to handle routine tasks inside Claude Code with a cheaper model; and Americas-based developers whose workdays fall almost entirely within Z.ai's discount window. Light users and those who want to connect their own applications are not a good fit.

Prompt Caching is the key to understanding this plan's economics. The principle is: when the prefix of multiple requests (system prompt, tool definitions, files already read, etc.) is exactly identical, the model provider can reuse the key-value cache (KV Cache) computed during the previous inference, without re-running forward computation on that portion of input. As a result, cache-hit input tokens enjoy an extremely low rate — in this article, GLM-5.3's fresh input is $1.40 per million tokens, while cached input is only $0.26 per million, a difference of about 5.4x.

Coding agents are naturally cache-dense scenarios: each loop step re-sends the entire conversation history, a single call might have 120,000 tokens of context, of which over 95% is a repeated prefix. Without caching, token-billed costs spiral out of control quickly; with caching, the cost of the same task can drop several-fold. This is also the fundamental reason the article repeatedly emphasizes "shorter context saves money" — context bloat not only increases input tokens but also lowers the cache hit rate.

2. 2026 Pricing and Tiers

Z.ai sells three individual tiers — Lite, Pro, and Max — all available monthly, quarterly, or annually. Monthly prices are $18, $80, and $168; quarterly gets 20% off, annual gets 30% off, with annual-equivalent monthly prices of $12.60, $56, and $117.60.

Tier	Monthly	Annual (monthly equiv.)	5-hour credits	Weekly credits
Lite	$18	$12.60	2,000	10,000
Pro	$80	$56	12,000	60,000
Max	$168	$117.60	28,000	140,000

The price per thousand weekly credits decreases with tier: Lite $1.80, Pro $1.33, Max $1.20. Pro's cost per credit is about 26% cheaper than Lite, and Max about a third cheaper. But this doesn't mean Max is the rational default — unused credits are worth zero. It means that once you're sure you'll use several times Lite's quota, the larger tier is the better buy: six Lite subscriptions at $108/month only buy the weekly credits Pro provides for $80.

Discounts, Referrals, and Team Pricing

New subscribers who sign up via an invite link get 10% off their first order, and the inviter receives credits worth 10% of the new subscriber's payment, released after 3 paying invitees. Team purchases are billed per seat: standard seat $88/seat/month, premium seat $188/seat/month, minimum two seats, with another 10% off for annual payment.

Worth noting is the price history. In April 2026, the monthly prices in Z.ai's migration notice were still $18, $72, and $160. Lite stayed at the same price, while Pro and Max rose by about 11% and 5% respectively when the credits system launched, and the usage metering method changed entirely. Any review written before July 30 is not describing the same product, even if the tier names are the same.

3. How Credits Work

Credits are directly tied to tokens, and the formula is public: credits = (input tokens × input multiplier + cached input × cache multiplier + output × output multiplier) / 10,000. MCP tool calls are billed separately at a fixed number of credits each.

Usage	Input multiplier	Cached input multiplier	Output multiplier
GLM-5.3	6.9	1.7	24
GLM-5.3-Flash	2.3	0.56	8
Search/Read/Zread MCP	—	—	1.2 per call

Two immediate consequences: cache hits are the lifeblood of this plan — when a coding agent re-sends about 120,000 tokens of context each step, the vast majority is billed at the 1.7 rate, which is what makes the advertised token quota possible. Reasoning is expensive — GLM-5.3's thinking tokens count as output tokens, billed at 24, and the model uses its highest reasoning effort by default.

Peak Hours and the 50% Discount

Outside peak hours, model usage is billed at 50% of the standard rate. Peak hours are Monday to Friday, 14:00 to 18:00 Singapore time (UTC+8), i.e., 20 hours out of 168 hours per week.

Region	Weekday peak window
Singapore, Beijing	14:00–18:00
India	11:30–15:30
Berlin, Paris	08:00–12:00
London	07:00–11:00
New York	02:00–06:00
San Francisco	Previous day 23:00–03:00

The pattern is clear: Americas-based developers almost never touch the peak window, effectively nearly doubling their quota; European developers pay full price for most of the morning; Indian developers pay full price at midday.

Two Windows

The 5-hour window refreshes on a rolling basis, with each usage released 5 hours after it's consumed; the weekly window resets every 7 days from the subscription moment. If either window is exhausted, tool calls fail, and the plan never falls back to account balance. Each tier's weekly quota is exactly 5x its 5-hour quota.

4. How Many Dollars Is the Quota Actually Worth

Z.ai's own conversion: at a 95% cache hit rate, Lite covers 48 million to 97 million GLM-5.3 tokens per week, Pro 290 million to 580 million, and Max 676 million to 1.352 billion. The low end assumes everything falls in peak hours, the high end assumes everything is off-peak.

Tier	Plan price	API value (all peak)	API value (all off-peak)
Lite	$18	$71	$141
Pro	$80	$424	$848
Max	$168	$989	$1,978

Read at Z.ai's own prices, the plan is a strong deal for heavy users: when fully used, Lite provides about 4 to 8x the monthly fee in usage, Pro 5 to 11x, Max 6 to 12x. But the more practical number is the break-even point — Lite needs about 53 million tokens purchased per month to be worthwhile, roughly a quarter of its peak quota; Pro about a fifth, Max about a sixth. Below that line, pay-as-you-go is cheaper.

The Flash Asymmetry

GLM-5.3-Flash paints an entirely different picture. The plan charges about a third of GLM-5.3's credits for it, but Z.ai's API prices Flash at about a ninth of GLM-5.3, so the plan's discount on Flash is much thinner. The practical rule is simple: spend your plan credits on GLM-5.3.

Comparing the Cheapest Hosts

The "save up to 92%" claim compares against Z.ai's own API, which is not the first procurement choice for budget-conscious developers. GLM-5.3's weights are open source, and OpenRouter hosts 32 providers. Under the same cache-dense workload, Z.ai's own endpoint ranks 31st out of 32.

Host	Input/Cache/Output (per million)	Cost per million tokens
Z.ai	$1.40 / $0.26 / $4.40	$0.34
DeepInfra	$0.56 / $0.13 / $2.50	$0.16
Novita	$0.42 / $0.08 / $1.32	$0.10
InferenceNet	$0.12 / $0.04 / $4.40	$0.07

Compared with the best hosts, the plan's advantage shrinks from "5 to 11x" to near break-even — Pro still wins at Novita's prices, but the multiple is about 1.6 rather than 5.

Open-weight models and the hosting ecosystem is an easily overlooked part of evaluating this plan's value. GLM-5.3's weights are public (open source), so any vendor can deploy it on its own GPUs and offer it at its own price. Aggregation platforms like OpenRouter put the same model from 32 providers side by side for price comparison, and the price differences can be enormous — InferenceNet's input price in the table is about 1/12 of Z.ai's official endpoint.

This has two consequences: first, Z.ai's advertised "save up to 92%" is compared against its own API, which happens to not be the cheapest procurement channel; second, the subscription plan's "fixed rate" is essentially ceding price volatility and vendor choice to Z.ai in exchange for predictable quota. If your workload is stable and your cache hit rate is high, going directly to third-party hosts like Novita or DeepInfra is often more flexible than subscribing — this article's calculations show the plan's advantage over the best hosts shrinks from over 5x to about 1.6x.

5. What Models You Actually Get

All tiers include only two models: GLM-5.3 and GLM-5.3-Flash. Routing rules determine what you actually get — requesting GLM-5.2 or GLM-5.1 is automatically routed to GLM-5.3, and requesting GLM-4.7 is routed to GLM-5.3-Flash.

GLM-5.3 uses the same base model as GLM-5.2, with all improvements coming from post-training, and features a 1 million token context, three levels of reasoning control, and open weights. It scores 28.3 on Terminal Bench 3.0 (up from 4.6 for GLM-5.2) and 66.9 on DeepSWE (versus 46.2 for the previous generation).

GLM-5.3-Flash, meanwhile, is a 320-billion-parameter, 18-billion-active natively multimodal model under MIT license, serving as the cheap workhorse within the plan.

The plan also includes four MCP tools: image understanding, web search, web reading, and Zread (GitHub repository retrieval), the latter three at 1.2 credits per call.

6. Configuration and Usage Tips

The core configuration requirement is: quota only takes effect when requests reach the Z.ai coding endpoint through supported tools. Each of the three protocols has its own base URL — Claude Code uses the Anthropic format (https://api.z.ai/api/anthropic), most editors use OpenAI Chat Completions (https://api.z.ai/api/coding/paas/v4), and Codex uses the Responses format (https://api.z.ai/api/v1).

Three Habits to Save Credits

There are three zero-cost habits to make your quota last longer: don't let context grow unbounded, lower reasoning effort for routine work, and move heavy tasks out of peak hours.

Keep context short: a single call with 120,000 tokens of context consumes 28.3 credits at peak, while 400,000 tokens costs 83 credits. In Claude Code, use /clear to start a new conversation, /compact to compress history, and /context to check window usage.

Lower the reasoning level: GLM-5.3 defaults to the highest reasoning effort, and every thinking token is billed at 24x output. In Claude Code, use /effort to switch levels — max for hard problems, high or low for mechanical work.

Leverage Flash and off-peak timing: hand small tasks to Flash (only a third of the credits), and schedule large refactors for off-peak hours to cut costs in half.

7. Rules That Could Cost You Your Subscription

The plan terms are stricter than most developer subscriptions, and enforced by automated "risk controls." The core restriction is that quota may only be used with officially supported tools, never for "general API access"; the terms explicitly list "SDK-based access" as a pattern the system can detect.

The second restriction concerns who may use it. The subscription is "licensed only to the natural person associated with the account," and sharing it in any way with colleagues, friends, clients, or organizations is prohibited. The actions most likely to trip up honest users are quite mundane: running an unlisted tool, wiring the Key into a script or internal bot, or lending the Key to a colleague for a day.

Subscriptions are non-refundable once purchased, and auto-renewal charges the price shown on the page on the renewal date — meaning price increases reach existing subscribers at their next renewal.

8. Should You Still Buy It?

The GLM Coding Plan is a good product, but its sweet spot is narrow. Three variables determine which side of the line you're on: how many hours per week your agent actually runs, how large your context grows, and whether those hours fall within Z.ai's peak window. Price matters less than these three — if the cheapest tier runs out by every afternoon, it's actually the expensive choice.

Your situation	Recommended purchase
A few tasks per week	Pay-as-you-go, prefer cheaper hosts
Daily use, in the Americas	Start with Lite, upgrade to Pro when you hit the wall
Daily use, in Europe or India	Pro, and move heavy work out of peak hours
Round-the-clock multi-project work	Max
Company team	Team seats with overage enabled

Whichever tier you pick, start with monthly billing for the first cycle and measure. The console shows credit consumption, and a week of real work at standard rates (not promotional periods) will tell you more than any estimate whether you're a Lite, Pro, or Max user.

Background Notes

Reasoning Effort / Thinking Budget refers to the token budget a model has for internal "thinking" before giving its final answer, typically divided into levels like low, high, and max. GLM-5.3 is a reasoning model, and the tokens generated by its thinking process count toward the output side, where the multiplier in the credits formula is as high as 24 (versus just 6.9 for input). This means that raising the reasoning level by one step can multiply your credit consumption, even if the final answer length stays the same.

The practical implication is: reasoning effort should be adjusted dynamically with task difficulty. Hard problems like architecture design and complex bug locating deserve max, letting the model fully unfold its chain of thought; mechanical work like formatting, simple rewriting, and boilerplate code generation should be lowered to high or low, saving both credits and response latency. In Claude Code you switch via the /effort command; most other tools set it in configuration options.

Endpoint here is not a technical detail but the compliance boundary for the subscription. Z.ai's coding plan only meters quota when traffic passes through specific supported tools and hits the dedicated coding endpoint; if you use the same API Key to directly request the general API endpoint, the system will identify it as "SDK-based access" and may trigger risk controls. In other words, the Key itself hasn't changed, but the protocol path taken determines whether this call is compliant plan usage or a violating general API call.

Common accidental triggers include: copying a Claude Code Key into your own Python script, connecting it to an internal Slack bot, or configuring the same Key in a third-party plugin outside Cline. These actions are technically feasible but bypass the plan's metering and authorization model, and fall under what the terms explicitly prohibit. Before use, confirm whether the tool is on Z.ai's official supported list.

One-line Summary

GLM Coding Plan offers three credit-based subscription tiers — best value for heavy Americas users, while light users save more with pay-as-you-go.

Paragraph Summary

The GLM Coding Plan is Z.ai's monthly subscription for coding agents, and since July 2026 it has switched from prompt counting to a token-metered credits system. The Lite, Pro, and Max tiers cost $18, $80, and $168 monthly, with 30% off annually, and the credits formula is public: input, cached input, and output each have multipliers, while MCP calls are billed at a fixed rate. Off-peak usage is billed at 50%, so Americas developers enjoy the discount almost all day, nearly doubling their effective quota. Plan credits are best spent on GLM-5.3, as the Flash discount is thin. Compared with Z.ai's own API you can save 4 to 12x, but against third-party open-source hosts like Novita the advantage shrinks to about 1.6x. Quota is limited to officially supported tools, and sharing or connecting it to your own applications is prohibited.