Table of Contents

• The table you came to see: GLM Coding Plan plans
• Models table: capabilities and API pricing

  • Text models
  • Vision, image, audio, and video models
• What almost no one explains to you: how credits are consumed
• The tools: where you can actually use this
• Use cases: which model for which task
• My real experience: three months, concrete flow
• The honest truth: who this is NOT for
• The invitation code: what each person gains, no frills
• Verdict

  • Sources

The table you came to see: GLM Coding Plan plans

This is the heart of Z.ai's offering for developers. All plans give access to the same models—no capped tiers—what changes is how much quota you have.

	Lite	Pro	Max
Base monthly price	~$18 USD	~$72–80 USD	~$160–168 USD
Credits every 5 hours	2,000	12,000	28,000
Weekly credits	10,000	60,000	140,000
Estimated tokens/week (GLM-5.3, cache ~95%)	48–97 million	290–580 million	676–1,352 million
Models included	GLM-5.3, GLM-5-Turbo, GLM-4.7	Same	Same
MCPs included	Vision, Web Search, Web Reader, Zread	Same	Same
Discount per cycle	20% quarterly / 30% annual	Same	Same

With the annual discount, Lite drops to ~$12.60/month. That's the number that makes the comparison with any other coding subscription uncomfortable for the competition.

Two price warnings: Pro and Max amounts have varied between sources and promotions ($72 vs $80, $160 vs $168), so verify the exact number on the page before paying. And subscriptions are not refundable once purchased—start with Lite monthly, not Max annual.

👉 Subscribe with 10% discount here — the discount only applies to your first payment and only if you've never had a paid subscription on the platform.

---

Models table: capabilities and API pricing

If you prefer to pay per token instead of subscribing, this is the catalog. Prices per million tokens (USD):

Text models

Model	Input	Input cached	Output	What it's for
GLM-5.3	$1.40	$0.26	$4.40	Flagship. Complex coding, long-term agentic tasks, security. Context 1M, output up to 128K
GLM-5.2	$1.40	$0.26	$4.40	Previous generation, same base as 5.3
GLM-5-Turbo	$1.20	$0.24	$4.00	Fast. Ideal for automations and low-latency responses
GLM-4.7	$0.60	$0.11	$2.20	Workhorse. Medium coding tasks
GLM-4.7-FlashX	$0.07	$0.01	$0.40	High volume: classification, extraction, routing
GLM-4.5-Air	$0.20	$0.03	$1.10	Lightweight, very cheap
GLM-4.7-Flash	Free	Free	Free	Prototypes and testing at no cost

Vision, image, audio, and video models

Model	Price	What it's for
GLM-5V-Turbo	$1.20 / $4.00 per MTok	High-level vision: UI, video, graphics analysis
GLM-4.6V	$0.30 / $0.90 per MTok	Standard vision, good cost-quality ratio
GLM-OCR	$0.03 per MTok	Text extraction from scanned documents
GLM-4.6V-Flash	Free	Basic vision at no cost
GLM-Image	$0.015 per image	Image generation
CogVideoX-3	$0.20 per video	Video generation
Vidu Q1 / Vidu 2	$0.20–0.40 per video	Video from text, image or reference
GLM-ASR-2512	~$0.0024 per minute	Audio transcription
Agent Slide/Poster	$0.70 per MTok	Automated presentations and posters

The most important data from this table: cached input costs one fifth of normal input. In an agentic flow where you resend the same project context over and over, that difference is what decides whether the month is expensive or cheap for you.

---

What almost no one explains to you: how credits are consumed

This is where people get frustrated and cancel. The plan doesn't work by "number of prompts". It works with two simultaneous limits and multipliers per model.

The two limits:

• 5-hour limit. Refreshes dynamically: each consumption is returned to you 5 hours later.
• Weekly limit. Starts when you pay and resets every 7 days.

The formula: credits = (tokens_input × mult_input + tokens_cached × mult_cache + tokens_output × mult_output) / 10,000

The multipliers:

Model	Input	Input cached	Output
GLM-5.3	6.9	1.7	24
GLM-5-Turbo	5.7	1.5	21
GLM-4.7	4.6	1.2	16
GLM-4.6V (Vision MCP)	1.2	0.3	2.7
Web Search / Web Reader / Zread	—	—	1.2 per call

Three practical consequences I learned the hard way:

1. Output is what's expensive. Multiplier 24 vs 6.9 for input in GLM-5.3. A model that "thinks out loud" for 40 thousand tokens will eat your day. That's why reasoning_effort: low exists and that's why I use it more than I expected.
2. Cache is your budget. The difference between 90.9% and 98% cache hit rate is the difference between 43 and 105 million tokens per week in Lite. Long, coherent sessions yield; opening new context every 10 minutes ruins you.
3. Off-peak hours cost half. Peak hours are Monday to Friday, 14:00–18:00 Singapore time (UTC+8). Everything else—including the entire weekend—costs 50% of credits. If you work from the Americas, a good part of your day falls in cheap hours. It's probably the least advertised price advantage of the plan.

---

The tools: where you can actually use this

The plan is strictly limited to officially supported tools. It's not an open API that you plug in wherever you want; if you use it outside the list, you'll get insufficient balance error.

Supported code agents: ZCode (Z.ai's own environment), Claude Code, Claude for IDE, Codex, Cursor, Cline, OpenCode, Roo Code, Kilo Code, Crush, Goose, Droid, TRAE, Qoder, Pi and Eigent.

General purpose agents: OpenClaw, Hermes Agent and SillyTavern, served with "best effort"—under high load they may apply temporary limits.

Configuration is changing the base URL:

Protocol	Base URL
Anthropic Messages (Claude Code, Goose)	https://api.z.ai/api/anthropic
OpenAI Chat Completions (the rest)	https://api.z.ai/api/coding/paas/v4
OpenAI Responses	https://api.z.ai/api/v1

And the four MCPs included at no extra cost beyond credits: Vision Understanding (read screenshots, diagrams, mockups), Web Search, Web Reader (extract content from URLs) and Zread (read GitHub repositories). The latter is more useful than it sounds when you're integrating a library you don't know.

---

Use cases: which model for which task

This is the part that separates those who squeeze the subscription from those who waste it.

GLM-5.3 → heavy and long work. Refactors that touch fifteen files, migrations, debugging something you've been chasing for hours, complete implementation of a spec. This is where the jump in this version is noticeable: in its internal benchmark it went from 23.4% to 34.5% completed tasks against GLM-5.2, and consuming fewer output tokens. In Terminal-Bench 3.0 it jumped from 4.6 to 28.3. Use it with reasoning_effort: max and don't use it for anything trivial.

GLM-5-Turbo → automations and latency. Everything that runs in a cron, webhook, or pipeline. Here you don't need deep reasoning, you need consistent and fast responses. It's also my default choice for chatbots where the user is waiting.

GLM-4.7 → daily work. Write an endpoint, adjust a test, explain a file I didn't write to you, generate a migration script. 16 multiplier on output vs 24 from GLM-5.3: it yields much more and for 60% of tasks the quality difference isn't noticeable.

GLM-4.7-FlashX / GLM-4.5-Air → volume. Classify tickets, extract text fields, route between agents, summarize. At $0.07 input per million, FlashX is practically free to process batches.

GLM-4.6V / GLM-5V-Turbo → vision. Pass it a UI bug screenshot, an architecture diagram in photo, a table in PDF. The Vision MCP integrated in the plan makes it especially comfortable within the coding flow.

A note about chatbots: if you're building conversational interaction with real users, the GLM-5-Turbo + Web Search MCP combination gives a noticeably more natural experience than the price would suggest. It's one of the uses where I've gotten the most value and the least talked about.

---

My real experience: three months, concrete flow

My way of working is this and I think it explains why it works so well for me:

I define specs with another model. I use Grok, GPT or Claude for the problem thinking part: architecture, tradeoffs, acceptance criteria, which files to touch. That conversation is exploratory, iterative, and there I want the best reasoning available regardless of cost, because it's few tokens.

I execute specs with GLM. Once the spec is written and unambiguous, GLM-5.3 implements it. And this is important: executing a well-defined spec is a different problem than designing it. It requires following instructions, maintaining context, not inventing, using tools correctly. GLM is very good at that, and that's where the price stops making sense in your favor.

Let's be clear, credits aren't enough to code all day in Lite. If your work is eight hours of agent spitting code, you'll hit the 5-hour limit. But for my pattern—bursts of execution interspersed with human review—it's plenty. And the 50% discount off-peak hours doubles my effective margin.

The app and chat also count. Outside the IDE I use Z.ai's app for tasks that don't require deep reasoning: review a topic, summarize documentation, verify something quickly, repetitive processes. Compared to opening a Sonnet for that, it's significantly more cost efficient and good enough. It's not where the magic is, but it's where the accumulated savings are.

---

The honest truth: who this is NOT for

If I'm clear about anything after three months, it's that this is not an entry-level tool.

It's NOT for you if:

• You're learning to program and need the model to guide you step by step, correct your assumptions, and explain to you why something is wrong. GLM executes very well; it won't rescue you from a poorly thought-out spec. If you give it vague instructions, it returns vague work with great confidence.
• You want an "all-in-one" experience without configuring anything. Here you need to understand base URLs, protocols, quota limits, multipliers and schedules. It's not difficult, but it's real friction.
• You need the absolute ceiling of capability. In Z.ai's own benchmark, GLM-5.3 reaches 34.5% where Claude Fable 5 reaches 39.5%. The gap exists and is felt in frontier tasks.
• You need a refund guarantee. There isn't one. Period.

It IS for you if: you already know how to define a problem, you already have criteria for reviewing what an agent produces, and your bottleneck is the cost of executing a lot, not the quality of reasoning at the frontier.

Two other things to consider before paying: Z.ai is a Chinese company (Zhipu AI) and your prompts go through their infrastructure—if you handle code under strict NDA or regulated data, review it with whoever is responsible. And plans are tied to supported tools: if your stack uses something not on that list, the subscription won't work for you and you'd have to go for paid API by token.

---

The invitation code: what each person gains, no frills

Since I'm going to share it, I prefer to explain exactly how the program works instead of just throwing you the link.

What you gain (guest):

• 10% instant discount on the amount of your first GLM Coding Plan subscription. It's applied automatically at checkout, without writing any code.
• Applies only if you're a new user or have never paid a subscription on the platform.
• Applies only to the first order. Renewals, upgrades and downgrades don't.
• Doesn't accumulate with other first purchase discount campaigns.
• You must complete payment within 72 hours of registering with the link.

What I gain (inviter):

• Platform credits equivalent to 10% of the amount you actually paid. It's not cash: it's credits that only serve to pay subscriptions, packs, or API calls within Z.ai. They don't cash out or transfer.
• And there's a threshold: I don't charge anything until I accumulate 3 valid guests. Before that, credits remain in pending status.

I'm telling you this in detail because the incentive changes how you should read a review, and I'd rather you know and decide for yourself.

👉 Link with 10% discount (code: V2TH8OSBEF)

Before using it: Lite monthly plan is ~$18 with the discount applied. Try it for a month with your real flow. If after 30 days you didn't hit the 5-hour limit and the model did what you needed, then yes, evaluate the annual with 30% off. Not the other way around.

---

Verdict

Z.ai is not "the cheap Claude". It's a tool with a different profile: excellent at executing well-defined work, at a price that changes the equation of how much you can automate. If your flow separates thinking from doing—defining specs with one model, executing them with another—it fits almost perfectly. If you expect one model to do everything, you'll get frustrated.

After three months my conclusion is simple: I feel it gives me more than I pay for. But that phrase only applies if you already know what to ask for.

---

Sources

• Z.AI Developer Docs — GLM Coding Plan Overview
• Z.AI Developer Docs — Pricing
• Z.AI Developer Docs — GLM-5.3
• Z.AI Developer Docs — Tool Integration
• Z.AI Developer Docs — GLM Coding Plan FAQ
• Z.AI Developer Docs — Invitation Program Rules
• TechNode — Z.ai launches GLM-5.3 with claimed 50% gain on coding benchmark
• VentureBeat — GLM-5.3 is here with advanced cyber capabilities
• MarkTechPost — Z.ai Ships GLM-5.3 Without Retraining the Base Model
• AI Pricing Guru — GLM Coding Plan Pricing

Disclosure: This article contains an invitation link. I receive platform credits if you subscribe through it. Technical and pricing data comes from Z.ai's official documentation and was verified in August 2026; verify current amounts before paying.