GLM-5.3 and Z.ai: What It Is, Pricing, Chat and Test Results

👁 127 reads

GLM is a family of language models from Chinese company Z.ai, formerly Zhipu AI. The latest flagship, GLM-5.3 from August 14, 2026, has 753 billion parameters, open weights, and costs $1.40 per million input tokens in the API. You can chat for free with the smaller GLM-5.3-Flash at chat.z.ai, and for coding there’s the GLM Coding Plan subscription starting at $18 per month.

What GLM Is and Who Z.ai Is

GLM (General Language Model) are language models developed since 2020 by Zhipu AI, a company spun out of the Knowledge Engineering Laboratory at Tsinghua University in Beijing. The company was founded on June 11, 2019 by professors Tang Jie and Li Juanzi. In July 2025, it adopted the name Z.ai for international markets, and on January 8, 2026, it went public in Hong Kong as the world’s first publicly traded large language model company. Since January 2025, it has been on the U.S. export restriction entity list.

Under the Z.ai brand, there are four products you ask about most often:

• chat.z.ai - free browser-based chat, running on GLM-5.3-Flash according to the site,
• Z.ai API - access to all GLM models, pay-per-token, compatible with OpenAI and Anthropic formats,
• GLM Coding Plan - subscription for coding in Claude Code, Cline, OpenCode and other tools,
• ZCode - Z.ai’s own coding application, which we covered in our article about 564 attempts to send code to a Chinese server.

We wrote about how Zhipu’s founder views open models in our article Zhipu founder wants open AI.

How to Get Started Step by Step

1. Simplest: open chat.z.ai and log in. The chat is free and responds in Polish. Remember this is GLM-5.3-Flash, not the full GLM-5.3.
2. For the full model via API: create an account at z.ai, add credit in USD, and create an API key. The OpenAI-compatible endpoint is https://api.z.ai/api/paas/v4/chat/completions, model glm-5.3.
3. Through OpenRouter: model z-ai/glm-5.3. If you care about consistent quality, pin the Z.AI provider - others serve the model in 4-bit quantization.
4. For coding: buy GLM Coding Plan and enter in ~/.claude/settings.json the address https://api.z.ai/api/anthropic and your key from the Z.ai console. Claude Code will then work on GLM-5.3. We described this step by step in our article GLM Coding Plan.
5. Set the reasoning level. GLM-5.3 always thinks, the reasoning_effort parameter accepts low, high, and max. For writing in Polish choose low, for difficult code max.

GLM Models

GLM-5.3 is built on the same base as GLM-5.2, with all progress coming from additional training. Z.ai held back the weights for two weeks for a safety review and published them in late August, but no longer under MIT: the GLM-5.3 license requires review by Z.ai for AI service providers with revenue above $10 billion. For regular companies and individuals, it’s practically an open model. The smaller GLM-5.3-Flash has a pure MIT license - before the official release it circulated on OpenRouter as anonymous ox-alpha.

On OpenRouter there’s also GLM-5.3 Prime (since September 23): the same model served faster, for double the price. In our September test of Chinese models, the Prime version on default reasoning thought to the end of the limit and didn’t return text, while on low level it wrote 993 words in 25 seconds.

Benchmarks

On launch day, Artificial Analysis gave GLM-5.3 60 points in its index, the same as Kimi K3 and 7 points more than GLM-5.2. After changing to index version 4.3, it has 45 points in the max variant and 34 points in low, and GLM-5.3-Flash has 42 points - all three at the top of open models. The biggest jump was in agentic tasks: in GDPval-AA the Elo rating rose from 1524 to 1770. Z.ai also boasts an 84.5% result in the CyberGym cybersecurity test, which we covered in our article GLM 5.3 for hacking. The hallucination rate in the knowledge test according to Artificial Analysis increased from 26% to 30%.

How Much GLM Costs

Pricing from docs.z.ai, prices in USD, October 3, 2026.

Z.ai documentation only lists the starting price of $18 - the full pricing is visible after logging in. External sources report $80 for Pro and $168 for Max per month, but we couldn’t verify this at the source. How credits are counted and how many tokens that represents, we detailed in our article GLM Coding Plan: pricing, limits and Claude Code. What GLM-5.3 really costs from different providers, we calculated in our article GLM 5.3: three numbers from the model card don’t add up.

GLM in Polish

In short: understands everything, writes with errors. It’s the weakest in Polish among the major Chinese models we’ve tested. In our September test of 13 writing tasks, GLM-5.3 scored 63.6 out of 100 (61.8 from Z.ai itself), with errors like “w ramach przeprosiny” (within apology) - details in our article GLM-5.3 in Polish. In our October test it was similar:

• in proofreading it left errors like “w każdym bądź razie“ (in any be case) and “bynajmniej” (by no means) instead of “przynajmniej” (at least) - it corrected 12 of 14 errors, other models 14 of 14,
• in translation it wrote awkward phrases mixing Polish and English words,
• in an email it met all requirements but started with an awkward construction.

However, its knowledge about Poland was flawless (20 out of 20), including the current president and the two capitals of Lubusz Voivodeship. For Polish texts, treat it as a quick draft that needs editing.

Our Test: GLM-5.3 vs Kimi, Qwen and DeepSeek

On October 3, 2026, we gave GLM-5.3 and four other Chinese models the same six tasks in Polish through OpenRouter, with low reasoning and provider pinned to Z.ai. We tested it in Polish. Full description in our article Kimi K3, GLM-5.3, Qwen3.8 Max and DeepSeek in Polish: test.

At low level, GLM didn’t reason at all in five of eight queries (0 reasoning tokens) and returned responses in 3-11 seconds. It wrote code in 4 seconds and passed all 29 tests. It made a mistake counting business days: it gave 21 instead of 20 for November 2026.

GLM vs Competition

Other guides: Kimi, Qwen, DeepSeek. Comparison with Western models on 13 tasks is in our article about lesser-known models from the top tier.

Who GLM Is For

• For developers who want a cheap Claude Code replacement. GLM Coding Plan from $18 and swapping the address in settings - that’s the simplest path to a top-tier open model.
• For quick API tasks: classification, data extraction, code. At low it responds in seconds for a fraction of a cent.
• For self-hosting - especially GLM-5.3-Flash under MIT license.
• Not for Polish texts without editing. And not for client code without checking where data goes - the team plan declares it doesn’t train on code, but we didn’t find such a declaration in the documentation for the individual plan.

Frequently Asked Questions

GLM - what is it?

GLM is a family of language models from Chinese company Z.ai (formerly Zhipu AI). The latest is GLM-5.3 from August 2026, with open weights and 1 million token context.

Are Z.ai and Zhipu the same company?

Yes. Zhipu AI adopted the international name Z.ai in July 2025, and since July 2026 the Hong Kong-listed company is called Z.AI Co., Ltd.

Is GLM free?

The chat.z.ai chat is free and runs on GLM-5.3-Flash. In the API, older Flash models are free (GLM-4.7-Flash, GLM-4.5-Flash, GLM-4.6V-Flash). Full GLM-5.3 is paid.

How much does GLM-5.3 cost?

In Z.ai API, $1.40 per million input tokens and $4.40 per million output. In GLM Coding Plan subscription from $18 per month.

Does GLM-5.3 work in Polish?

It understands Polish well but writes with errors: in our test it corrected 12 of 14 errors in proofreading and translated idioms the weakest. Its knowledge about Poland was flawless.

How to use GLM in Claude Code?

Buy GLM Coding Plan, create a key in the Z.ai console, and set in ~/.claude/settings.json the variables ANTHROPIC_BASE_URL to https://api.z.ai/api/anthropic and ANTHROPIC_AUTH_TOKEN to your key.

Is GLM-5.3 open source?

Weights are open on Hugging Face, but under the GLM-5.3 license, not MIT. The restriction only applies to AI service providers with revenue above $10 billion. GLM-5.3-Flash has an MIT license.

GLM or DeepSeek?

For Polish texts DeepSeek: in our test 93.3 points vs 80.1 for GLM. For quick tasks and code in subscription GLM, because it was fastest in the test and has a ready plan for Claude Code.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and Verification Date

API pricing

• docs.z.ai pricing. GLM Coding Plan: docs.z.ai documentation (plan overview
• team plan
• Claude Code configuration
• plan change announcement from July 30
• 2026). Model in chat: chat.z.ai. Parameters and license: model cards on Hugging Face (zai-org/GLM-5.3
• zai-org/GLM-5.3-Flash). Company history and IPO: Wikipedia (Z.ai
• Zhipu AI)
• CNBC and Caixin. Index: Artificial Analysis. Providers: OpenRouter. Test results are our measurements through OpenRouter. Verified October 3
• 2026

Kinetyra