GLM-5.3
Z.ai's post-trained successor to GLM-5.2, much stronger at long-horizon coding and cyber tasks
Context window
1M
Input / 1M tokens
$1.40
Output / 1M tokens
$4.40
Provider
Z.ai (Zhipu AI)
Z.ai API list price; cached input $0.26 per 1M, with cached-input storage free for a limited time. Same price as GLM-5.2. Open weights are free to self-host under the GLM-5.3 License, which requires companies running a model-as-a-service business with more than $10B in yearly revenue to pass a Z.ai security review before commercial use. · Data verified 2026-11-08
GLM-5.3 (August 14, 2026) is Z.ai's flagship text model. It uses the same base model as GLM-5.2, and all of its gains come from post-training aimed at complex coding and long-horizon agent tasks. It has a 1M-token context window, up to 128K output tokens and always-on reasoning with low, high and max effort levels. Z.ai reports large jumps over GLM-5.2, such as 28.3 vs 4.6 on Terminal-Bench 3.0 and 66.9% vs 46.2% on DeepSWE v1.1, along with much stronger cybersecurity results. Open weights followed on Hugging Face later in August under a custom GLM-5.3 License rather than MIT.
How to access it
Z.ai API (model glm-5.3), the GLM Coding Plan, OpenRouter (z-ai/glm-5.3) and other hosts including Amazon Bedrock, Together, Fireworks and Alibaba QwenCloud. Open weights are on Hugging Face (zai-org/GLM-5.3) for vLLM, SGLang and other engines.
Strengths
- ✓Large post-training gains over GLM-5.2: 28.3 vs 4.6 on Terminal-Bench 3.0 and 66.9% vs 46.2% on DeepSWE v1.1 (Z.ai-reported)
- ✓1M-token context with up to 128K output tokens
- ✓Same $1.40/$4.40 per 1M price as GLM-5.2
- ✓Strong cybersecurity results, such as 84.5% on CyberGym (Z.ai-reported)
- ✓Open weights, plus hosted versions from many providers including Amazon Bedrock, Together and Fireworks
Best for developers who...
When to choose it (and when not to)
Reach for GLM-5.3 when...
- →You want Z.ai's strongest coding and agent model at $1.40/$4.40 per 1M
- →You run long agent sessions that benefit from a 1M-token context
- →You need an open-weight model for defensive security research
- →You already use GLM-5.2: same price, same base model, better post-training
Look elsewhere if...
- ✕You need image or video input (GLM-5.3 is text-only; GLM-5.3-Flash is multimodal)
- ✕You want a plain MIT license (GLM-5.2 and GLM-5.3-Flash are MIT)
- ✕You need to switch reasoning off: it is always on
- ✕Cost per token matters most (GLM-5.3-Flash costs $0.15/$0.50)
How to use it
- ›Reasoning is always on; set reasoning_effort to low, high or max (max is the default)
- ›Z.ai's examples use temperature 1.0
- ›Reuse long repository context across calls: cached input costs $0.26 per 1M instead of $1.40
Quickstart
Pythonfrom openai import OpenAI
client = OpenAI(
api_key="YOUR_ZAI_API_KEY",
base_url="https://api.z.ai/api/paas/v4/",
)
resp = client.chat.completions.create(
model="glm-5.3",
messages=[{"role": "user", "content": "Find the bug in this function and fix it."}],
extra_body={"thinking": {"type": "enabled"}, "reasoning_effort": "max"},
)
print(resp.choices[0].message.content)Base URL and parameters from Z.ai's GLM-5.3 docs. Self-host with: vllm serve zai-org/GLM-5.3
API model id: glm-5.3
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Terminal-Bench 2.1 | 88.2 | Z.ai-reported; GLM-5.2 81.0 |
| Terminal-Bench 3.0 | 28.3 | Z.ai-reported; GLM-5.2 4.6 |
| DeepSWE v1.1 | 66.9% | Z.ai-reported; GLM-5.2 46.2% |
| CyberGym | 84.5% | Z.ai-reported; GLM-5.2 77.2% |
| Toolathlon Verified | 73.0 | Z.ai-reported; GLM-5.2 59.9 |
| HLE (with tools) | 62.5 | Z.ai-reported; GLM-5.2 54.7 |
| GDPval-AA v2 | 1769 | Z.ai-reported; GLM-5.2 1508 |
Compare GLM-5.3
Compare GLM-5.3 with any other model
Build a comparison →All model comparisons →