Skip to content
freellm.site

Zhipu AI Explained: GLM Models, Pricing, Free Tier

Zhipu AI is now Z.ai, the Tsinghua-born lab behind the GLM models. Here's the 2026 model lineup, pricing, free tier and how to call the API.

By
zhipu-aiz.aiglmopen-source-llmfree-llm-apichinese-ai

Zhipu AI is the Chinese foundation-model lab spun out of Tsinghua University’s Knowledge Engineering Group. It builds the GLM family of open-weight models, currently led by GLM-5.3 and GLM-5.3-Flash. It rebranded internationally as Z.ai in July 2025 and listed on the Hong Kong Stock Exchange on 8 January 2026, pricing shares at HK$116.20 for a valuation near HK$51 billion. If you searched “zhipu ai,” you probably want to know what the company is, which GLM model to use, or how to get an API key. This post covers all three.

Zhipu AI Explained: GLM Models, Pricing, Free Tier

The naming trips people up first. Zhipu AI and Z.ai are the same company. Older model cards, papers and blog posts say Zhipu; the current docs, API domain (api.z.ai) and Hugging Face org (zai-org) all say Z.ai. The GLM model line continues across both names, so a 2025 “Zhipu GLM-4.6” reference and a 2026 “Z.ai GLM-5.3” reference belong to the same lineage.

What Zhipu AI ships

It ships four separate products, which is why advice about “the Zhipu API” often contradicts itself:

  1. Open weights. GLM models are published on Hugging Face and ModelScope and run locally. GLM-5.3-Flash weights are on Hugging Face with support for SGLang, vLLM and TokenSpeed. For GLM-5.3, Z.ai said weights would follow two weeks after launch, once safety evaluation and hardening were done.
  2. The pay-as-you-go API at https://api.z.ai/api/paas/v4/, which is OpenAI-compatible.
  3. The GLM Coding Plan, a subscription that routes GLM models into Claude Code, ZCode, Cline, OpenCode and 20+ other agent tools. It starts at $18/month per the DevPack overview.
  4. Free-tier models. Several Flash-class models are listed at $0 on the pricing page (details below).

Because the weights are open, GLM models also show up on third-party hosts. If you’re comparing routes, our roundup of the best open source LLM APIs in 2026 covers who serves what.

The 2026 GLM lineup

GLM-5.3: the flagship

GLM-5.3 launched 14 August 2026. It uses the same base model as GLM-5.2, so every gain came from scaled post-training: more environments, more task diversity, more compute. Z.ai reports a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench, plus open-source SOTA on Terminal Bench 3.0 and Agents’ Last Exam (CLI).

The release also documented a result the team didn’t plan for: cyber capability improved faster than expected after vulnerability-discovery data entered the training mix. GLM-5.3 is state of the art on CyberGym for vulnerability discovery and more than doubles GLM-5.2 on exploitation benchmarks.

GLM-5.3 is text-only, with a 1M-token context window and 128K max output. It always runs with reasoning enabled; you choose a reasoning_effort of low, high or max (default max). Per the GLM-5.3 model guide, if your app sends thinking.type: "disabled", you must switch it to enabled with reasoning_effort: "low" before pointing it at glm-5.3, or the request fails.

GLM-5.3-Flash: the cheap multimodal one

GLM-5.3-Flash arrived 26 August 2026 as the first natively multimodal model in the GLM-5 series, with 320B total parameters and 18B active. Its architecture is new for the series: a hybrid of sparse and linear attention plus Manifold-Constrained Hyper-Connections (mHC), trained on a 30T-token multimodal corpus. Z.ai reports it uses 3.0× less attention compute and a 4.4× smaller KV cache than GLM-5.3.

It scores 57 on the Artificial Analysis Intelligence Index v4.1.1 at $0.045 per task (discounted), which Z.ai pitches as near-frontier intelligence at Flash pricing. Before launch it was A/B tested anonymously as ox-alpha on OpenCode and OpenRouter, and all of that traffic ran on Chinese AI chips.

GLM-5.3-FlashX and the back catalogue

GLM-5.3-FlashX is the throughput tier, advertised at 200 tokens/s. There’s also a GLM-5.3-FlashX-Legal variant for legal research. Older generations (GLM-5.2, GLM-5.1, GLM-5, GLM-4.7, GLM-4.6) can still be called, though the Coding Plan auto-routes GLM-5.2/5.1 requests to GLM-5.3 and GLM-4.7 to GLM-5.3-Flash.

Zhipu AI pricing (per 1M tokens)

From the official Z.ai pricing page, in USD. Cached-input storage was listed as “limited-time free” when we wrote this, so re-check before you budget around it.

ModelInputCached inputOutputContext
GLM-5.3$1.40$0.26$4.401M
GLM-5.3-FlashX$0.37$0.075$1.251M
GLM-5.3-Flash$0.15$0.03$0.501M
GLM-4.7$0.60$0.11$2.20200K
GLM-4.6$0.60$0.11$2.20200K
GLM-4.7-FlashX$0.07$0.01$0.40check docs
GLM-4.7-FlashFreeFreeFreecheck docs
GLM-4.5-FlashFreeFreeFreecheck docs
GLM-4.6V-Flash (vision)FreeFreeFreecheck docs

Non-text pricing: GLM-Image at $0.015/image, CogView-4 at $0.01/image, CogVideoX-3 at $0.2/video, GLM-ASR-2512 at $0.03/MTok, and the built-in Web Search tool at $0.01 per use. The pricing page doesn’t publish rate limits for the free models. Our Zhipu AI free API provider page lists what we’ve verified, and the official docs have the authoritative numbers.

Benchmarks as reported by Z.ai

These are the vendor’s own numbers from the GLM-5.3 announcement, not an independent audit. We’ve written about how easily benchmark charts mislead.

BenchmarkGLM-5.3GLM-5.2Opus 4.8
Terminal Bench 3.028.34.621.1
DeepSWE v1.166.946.258.0
CyberGym84.577.278.1
Toolathlon Verified73.059.976.2
HLE w/ Tools62.554.757.9

The token-efficiency numbers may matter more than the raw scores. At High effort, Z.ai reports GLM-5.3 reaching 31.4% on its agentic coding eval with ~50K output tokens, against Opus 4.8 at 29.5% with ~120K. Z.ai also says GLM-5.3 still trails Claude Fable 5 (39.5% at Max effort). Vendors rarely name the model that beat them.

How to call the Zhipu AI / Z.ai API

Get a key from the Z.ai Open Platform, then use any OpenAI-compatible client. The base URL is https://api.z.ai/api/paas/v4/:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["ZAI_API_KEY"],
    base_url="https://api.z.ai/api/paas/v4/",
)

resp = client.chat.completions.create(
    model="glm-5.3",
    messages=[
        {"role": "system", "content": "You are a senior backend engineer."},
        {"role": "user", "content": "Refactor this Flask route to use async SQLAlchemy."},
    ],
    extra_body={
        "thinking": {"type": "enabled"},
        "reasoning_effort": "max",   # low | high | max
    },
)

print(resp.choices[0].message.content)

Swap glm-5.3 for glm-4.7-flash to test on the free tier first. If this pattern is new to you, what an OpenAI-compatible API is explains why the same few lines work across a dozen providers.

Raw HTTP works the same way:

curl https://api.z.ai/api/paas/v4/chat/completions \
  -H "Authorization: Bearer $ZAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"glm-5.3-flash","messages":[{"role":"user","content":"Hello"}]}'

Besides plain API-key auth, the docs describe a JWT authentication path for higher-security deployments.

Using GLM inside Claude Code

The Coding Plan maps GLM models onto Claude Code’s model slots through environment variables. By default ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL and ANTHROPIC_DEFAULT_HAIKU_MODEL all point at GLM-5.3-Flash. Z.ai ships a helper to set it up:

npx @z_ai/coding-helper

It installs supported tools, loads your plan credentials and manages MCP servers (Web Search, Web Reader, Zread). The Claude Code setup guide has the manual steps if you’d rather edit ~/.claude/settings.json yourself. Our walkthrough on using a free LLM API in Cursor or Claude Code applies the same substitution to other providers.

How Zhipu AI compares

GLM is an open-weight coding and agent line that competes on price per unit of capability and token efficiency. It doesn’t top every leaderboard. Against the open-weight models Z.ai benchmarks itself against (Kimi K3, DeepSeek-V4, Qwen3.8-Max), GLM-5.3 wins some tasks and loses others: it leads on Terminal Bench 3.0 and CyberGym and trails Kimi K3 on SWE-Marathon.

Against closed frontier models, the argument is cost. GLM-5.3-Flash at $0.15 input is roughly an order of magnitude cheaper than closed flagship pricing, and Z.ai’s Artificial Analysis figures put it in a competitive intelligence band. For where open weights stood this year, read the state of AI in 2026, and Muse Spark Open covers the other big open-weight release.

On free tiers, GLM’s permanently free Flash models put it alongside the providers in our best free LLM APIs in 2026 list. GLM-4.6V-Flash also gives you free vision, which most free tiers don’t.

FAQ

Is Zhipu AI the same as Z.ai?

Yes. Zhipu AI rebranded internationally as Z.ai in July 2025 and is listed in Hong Kong as Z.AI Co., Ltd. The GLM model family is the same under both names, and the current docs, API domain and Hugging Face org all use Z.ai.

Does Zhipu AI have a free API?

Yes. The official pricing page lists GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash (vision) at $0 for input, cached input and output. Rate limits and quotas aren’t on the pricing page, so check the official docs or our Zhipu AI provider page before building on them.

Are GLM models open source?

The weights are published. GLM-5.3-Flash is on Hugging Face under the zai-org org with SGLang, vLLM and TokenSpeed support, and GLM-5.3 weights were scheduled for two weeks after launch, pending safety hardening. Licences differ between releases, so read each model card.

Which GLM model should I use for coding?

GLM-5.3 at reasoning_effort: "max" is the flagship for complex, long-running engineering work. GLM-5.3-Flash is the value pick: cheaper, multimodal, and what the Coding Plan maps onto Claude Code’s slots by default. GLM-5.3-FlashX targets throughput at a stated 200 tokens/s. Our best free LLM APIs for coding guide compares these with non-GLM options.

What is the GLM Coding Plan and is it worth it over the API?

It’s a subscription (from $18/month) that works with Claude Code, ZCode, Cline, OpenCode and 20+ other agent tools, with higher call limits than pay-as-you-go and off-peak discounts. Z.ai claims savings of up to 92% over standard pay-as-you-go GLM-5.3 calls if you make full use of off-peak windows. Whether it beats metered billing depends on your cache-hit rate and usage pattern; the DevPack docs publish estimated weekly token allowances for each tier.