Z.ai and its GLM models: which is which and when to use each one

Table of Contents

• The map in 30 seconds
• The thinkers: text models

  • GLM-5.3 — the flagship
  • GLM-5.3-Flash — the workhorse
  • GLM-5.2 — the MIT option
  • GLM-4.7, 4.6 and friends — the generation that stayed
• The seers: vision and documents

  • GLM-4.6V — the one that understands what it sees
  • GLM-OCR — the cheap specialist
• The makers: image, video and voice
• How to use them correctly: five rules

  • 1. Pick by task, not by version number
  • 2. Use the right channel for each thing
  • 3. Context is not free (nor harmless)
  • 4. Specialists before all-rounders
  • 5. Don't marry the exotic variants
• Quick decision guide
• What actually changes between Coding Plan tiers

The map in 30 seconds

Model	Type	What it's for	API price
GLM-5.3	Text	The flagship: hard code, agents, long tasks	$1.40 / $4.40
GLM-5.3-Flash	Text + image	Volume: chat, extraction, classification	$0.15 / $0.50
GLM-5.2	Text	The MIT alternative with lossless 1M context	$1.40 / $4.40
GLM-4.7 / 4.6	Text	The previous generation, still solid all-round	$0.60 / $2.20
GLM-4.7-Flash	Text	Free on the API: sketches and side projects	$0
GLM-4.6V	Vision	Screenshots, documents, UIs, native tool use	$0.30 / $0.90
GLM-4.6V-Flash	Vision	Same idea, free	$0
GLM-OCR	OCR	Scans and PDFs into clean text	$0.03 / $0.03
GLM-Image	Image	Generation with readable text inside the image	~$0.015 per image
CogVideoX-3	Video	Short clips	~$0.20 per video
GLM-ASR-2512	Voice	Audio transcription	~$0.0024 per minute

If you take only one idea from this table: the "type" column matters more than the version number. A text 5.3 won't help you read a screenshot, and a 4.6V shouldn't be used for what GLM-OCR does at a tenth of the price.

---

The thinkers: text models

GLM-5.3 — the flagship

It's the top model of the family (August 2026) and the one that answers by default when you ask for "the best". It's optimized for code and agents: on Z.ai's own benchmarks it improved ~50% over GLM-5.2, and on terminal tasks and long refactors it sets the open-source state of the art. 1-million-token context and up to 128K of output — enough for an agent to write a whole feature in a single run.

One pricing detail almost nobody looks at: cached input costs $0.26 instead of $1.40. If your app sends the same system prompt on every call, that's nearly an 80% automatic discount on the repeated part.

Use it for: complex refactors, autonomous agents, large codebase analysis, hard reasoning. Don't use it for: classifying tickets or summarizing emails. That's Flash territory, and the bill difference is 10x.

GLM-5.3-Flash — the workhorse

The budget version (August 2026): 320B total parameters, 18B active, a hybrid architecture that cuts cache and compute costs. It accepts images in addition to text, also with 1M context, and it's the default model in Z.ai's chat.

At $0.15/$0.50 per million tokens, it's among the cheapest options in its class on the market. Rule of thumb: if the task bores you, it's for Flash; if it costs you effort, it's for full 5.3.

Use it for: chatbots, structured extraction, classification, summaries, anything at volume. Don't use it for: software architecture or decisions production depends on. There, the flagship's quality points pay for themselves.

GLM-5.2 — the MIT option

June 2026, same price as 5.3 on the API. Why does it exist, then? The license: GLM-5.3 shipped weights under its own license (with a security review required for giant AI-service companies), while GLM-5.2 is pure MIT. It was flagged as probably the most capable open-weight model at launch, with lossless 1M context (less goal-forgetting on long tasks).

Use it for: self-hosting in a company without talking to lawyers, or when your policy only allows MIT/Apache licenses.

GLM-4.7, 4.6 and friends — the generation that stayed

GLM-4.7 (late 2025) is still a serious model — the last with a public SWE-bench Verified score, 73.8% — at $0.60/$2.20. GLM-4.6 is roughly the same price. And GLM-4.7-Flash is free on the API: 30B total parameters, 3B active, perfect for prototypes, webhooks and automations that barely reason.

The GLM-4.5 and 4.5-Air from mid-2025 opened this era, but today they're history: 128K context and less capability than any of the above. If an old tutorial tells you "use GLM-4.5", swap it for 4.6 or newer.

---

The seers: vision and documents

GLM-4.6V — the one that understands what it sees

The vision model (106B, 128K context): it understands screenshots, documents with layout, product photos, UIs. Its trick is native tool use: it can look at an interface and chain actions on it, from visual perception to online action. At $0.30/$0.90 it's notably cheaper than the text flagships.

Use it for: agents that navigate interfaces, extraction where document layout matters, detailed image description. The trap: if you only need the text from a scanned PDF, GLM-OCR does that job at $0.03.

GLM-OCR — the cheap specialist

A tiny model (0.9B) that does exactly one thing: turn scans into clean text. The correct pattern is using it as the first step of a pipeline: OCR first, text model after. Processing 500 invoices this way costs cents; feeding them all to the flagship with image input costs dollars.

---

The makers: image, video and voice

• GLM-Image (January 2026): image generation at ~$0.015 per unit, with an unusual strong point: text inside the image comes out readable (signs, covers, mockups). MIT open weights.
• CogView-4: the house's historic generator, at ~$0.01 per image. Simpler, cheaper.
• CogVideoX-3: short video at ~$0.20 per clip. For ad drafts and social media material, not final production.
• GLM-ASR-2512: speech-to-text at ~$0.0024 per minute of audio. Ridiculously cheap for subtitling.
• GLM-TTS: the inverse, text to speech, with open weights.
• AutoGLM-Phone: the experimental 9B agent that operates a phone. A curiosity today, a signal of where the house is heading.

---

How to use them correctly: five rules

1. Pick by task, not by version number

The most common mistake is treating the catalog as a ladder where "newest is best". Flash isn't "the bad GLM": it's the GLM for simple tasks at volume, and on those tasks it matches quality at a tenth of the cost. The correct selection is horizontal: first the type (text, vision, OCR, image), then the tier (flagship or Flash), and only at the end the version.

2. Use the right channel for each thing

Z.ai has three doors, and mixing them is the number one source of surprise bills:

• chat.z.ai — for trying and chatting, free. Ideal before writing a single line of code.
• Pay-as-you-go API — for your apps. OpenAI-compatible endpoint:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://api.z.ai/api/paas/v4",
)

response = client.chat.completions.create(
    model="glm-5.3-flash",
    messages=[{"role": "user", "content": "Extract name and total from this invoice"}],
)

• GLM Coding Plan — the subscription for coding tools (Claude Code, Cline, OpenCode, Roo and friends). Here's the detail that burns: the plan uses its own coding endpoint and its own API key. If you configure the pay-as-you-go key on the generic endpoint, it doesn't deduct the plan's quota — it bills separately. The full walkthrough is in the onboarding post.

3. Context is not free (nor harmless)

With a 1M-token window, the temptation is to paste the whole repository "because it fits". Two problems: you pay for every token you send, and models degrade or forget the goal when the prompt is a haystack. Long context replaces badly-done RAG, not well-done RAG. And turn on prompt caching: the 5.3's cached input costs $0.26 instead of $1.40 — if your system prompt is fixed, that difference is pure money.

4. Specialists before all-rounders

The pattern repeats across the whole catalog: the cheap pipeline beats the expensive universal model. Scanned PDF → GLM-OCR ($0.03) + Flash, not the flagship staring at images. UI screenshot → 4.6V, not the text one. Blog image → GLM-Image ($0.015), not a premium tool. Every specialist exists because it does one thing better or a hundred times cheaper than the generalist.

5. Don't marry the exotic variants

In August 2026, Z.ai quietly pulled the GLM-5-Turbo and GLM-5V-Turbo variants from the API. Base models and Flash lines have been stable for years; experiments get retired. If your production hardcodes a model name, make it a mainline one — and keep a fallback configurable via environment variable, not via commit.

---

Quick decision guide

• Coding with agents (Claude Code, Cline, OpenCode) → GLM Coding Plan + GLM-5.3. Step-by-step activation here.
• Everyday chat, studying, drafts → chat.z.ai free (5.3-Flash by default).
• App with a volume of simple calls → GLM-5.3-Flash on the API.
• Hard reasoning in production → GLM-5.3.
• Self-hosting without license discussions → GLM-5.2 (MIT), or the Flash models on small hardware (Ollama guide here).
• Sketches and trivial automations → GLM-4.7-Flash, free on the API.
• PDFs and scans → GLM-OCR + Flash; if layout matters, GLM-4.6V.
• Understanding screenshots, UIs, visual documents → GLM-4.6V.
• Images → GLM-Image (readable text) or CogView-4 (cheaper).
• Short video → CogVideoX-3.
• Transcriptions → GLM-ASR-2512.
• Data that can't leave your machine → MIT weights on your own hardware.

What actually changes between Coding Plan tiers

If your case is coding, the subscription has three tiers: Lite ($18/month, 10,000 weekly credits), Pro ($80/month, 60,000) and Max ($168/month, 140,000). All three use the same models — 5.3 included; what changes is quota. Two little-known multipliers: paying quarterly or annually takes 20% and 30% off, and requests outside peak hours (outside 14:00–18:00 Singapore time on weekdays) consume half the credits. Coding at night in Latin America is, quite literally, half price.

The deep comparison against OpenCode Go — when each subscription makes sense — is in this post, and the real three-month usage experience, in the review.

---

Prices and features cited are as of September 2026 and change frequently. Verify at docs.z.ai before making any serious decision.

Want to try it? With this invitation link you get 10% off your first Coding Plan order: you save on the first payment and I receive platform credit.

— Leandro Valencia