GLM-5.3

glm-5.3

GLM-5.3 is Zhipu AI's latest flagship model, with comprehensive advancements in complex software engineering and Agent tasks. It uses the same base model as GLM-5.2—all improvements come from post-training. Compared with GLM-5.2, it performs even better on complex programming and long-range tasks.

Context window1.0M

Provider智谱AI

Released2026/09/01

Playground

Pricing

The same model is available through multiple service channels — choose based on latency, reliability and cost.

Prices in $ / 1M tokens

To pick a channel, add a provider field to the request body, for example "provider": { "channel": "direct" }. Valid values are direct / stable / economical; omit it to use the default channel.

Direct

Direct upstream connection — best when you need native behavior and the full context window.

Input	Output	Cache read
1.25/M	4.40/M	0.30/M

Overview

Input

Text

Output

Text

GLM-5.3 API: The Open-Weight Model with the Strongest Coding Capabilities

GLM-5.3 is the flagship model released by Z.ai on August 14, 2026, with weights opened on August 28. It doesn't change the base architecture—it continues to use GLM-5.2's MoE base (753 billion total parameters, ~40 billion activated per token), with all improvements coming from post-training scaling: a tenfold increase in long-horizon task environments, richer environment types, and significantly extended post-training time.

The result is a 50% improvement in coding capability over GLM-5.2 on Z.ai's self-developed Z.ai Code Bench, and the top ranking among open-weight models on public benchmarks such as Terminal-Bench 3.0 and Agents' Last Exam. Another direction the official team didn't anticipate was cybersecurity: CyberGym 84.5, ExploitBench 54.4, with vulnerability discovery and white-box code audit capabilities significantly exceeding expectations.

SeaWhale AI provides GLM-5.3 through an OpenAI-compatible interface, supporting tool calls, reasoning_effort reasoning intensity control, and streaming output.

Get API Key · Model ID: GLM-5.3

Why Choose GLM-5.3

Strongest coding among open-weight models — 50% improvement over GLM-5.2 on Z.ai Code Bench
#1 among open models on agent benchmarks — Terminal-Bench 3.0 jumped from 4.6 to 28.3, Agents' Last Exam 28.5
Outstanding cybersecurity capabilities — CyberGym 84.5, ExploitBench 54.4, with white-box audit and vulnerability discovery as key strengths
1 million token context — maximum output of 128K tokens
Shorter execution paths — uses more than half fewer output tokens than closed-source flagships for similar tasks

Core Capabilities

01 Project-Level Coding and Engineering Agent

This generation focuses post-training on long-horizon task environments, with cross-file feature implementation, refactoring, and debugging all significantly more stable than GLM-5.2. On Z.ai Code Bench's High tier, GLM-5.3 achieves 31.4% accuracy with about 50K tokens, while Claude Opus 4.8 needs about 120K tokens at the highest tier to reach 29.5%—for the same task, GLM-5.3's execution path is significantly shorter.

Cross-file, cross-module feature development and refactoring
Long-chain defect localization and fixing
Agent-style autonomous engineering workflows

02 Cybersecurity and Vulnerability Discovery

The official team describes this capability as "developing faster than expected." In white-box code audit, vulnerability discovery, and exploit chain construction, GLM-5.3 is in the top tier of current open-weight models, with CyberGym 84.5 being the SOTA for this benchmark.

White-box code security audit
Vulnerability discovery and validation
Automated security engineering workflows

03 1 Million Token Long Context

With a 1 million token context and 128K maximum output, it can read an entire medium-to-large code repository, hundreds of pages of PDFs, or a complete long conversation history in one pass, then output a complete analysis or change plan.

Complete code repository-level analysis
Extra-long technical document processing
Full preservation of long conversation history

04 Three-Level Reasoning Intensity

GLM-5.3's thinking mode is always on, providing low, high, and max levels via reasoning_effort (default max). Drop to a lower level for simple tasks to save tokens, or use max for complex engineering tasks to get the most complete reasoning chain.

low — lightweight tasks, cost control
high — routine development and analysis
max — complex engineering and agent tasks (default)

Best Use Cases

Scenario	Description
Project-level coding	Coding SOTA among open-weight models, short execution paths
Coding agents	#1 among open models on Terminal-Bench 3.0, Agents' Last Exam
Security audit	White-box code audit and vulnerability discovery, CyberGym SOTA
Long-context analysis	One-pass processing of entire repositories, hundreds of pages of documents
Domestic replacement	Open weights, can be privately deployed
Cost optimization	Significantly fewer output tokens for equivalent results

Differences Between GLM-5.3, GLM-5.2, and GLM-5.3-Flash

Capability	GLM-5.3	GLM-5.2	GLM-5.3-Flash
Model ID	`GLM-5.3`	`GLM-5.2`	`GLM-5.3-Flash`
Total params / activated	753B / ~40B	Same base	320B / 18B
Context window	1M tokens	1M tokens	1M tokens
Max output	128K tokens	128K tokens	128K tokens
Input modality	Text	Text	Text + Image
Terminal-Bench 3.0	28.3 (#1 open)	4.6	Below GLM-5.3
Weight license	GLM-5.3 License (commercial use allowed)	MIT	MIT
Reasoning intensity	low / high / max	high / xhigh	low / high / max
Positioning	Flagship: strongest coding and security	Previous-gen flagship	Lightweight multimodal, ~1/10 price

Specific billing is subject to the real-time price card at the top of the page.

FAQ

When was GLM-5.3 released? Released on August 14, 2026, API launched on August 19, and model weights opened on August 28.

What improvements does it offer over GLM-5.2? The base model hasn't changed; all improvements come from post-training scaling. Coding capability improved 50% on Z.ai Code Bench, Terminal-Bench 3.0 went from 4.6 to 28.3, Agents' Last Exam went from 23.8 to 28.5, and a new strong cybersecurity capability emerged.

How does it compare to closed-source flagships? Coding and agent capabilities are close to Claude Fable 5. On Terminal-Bench 3.0, 28.3 still trails GPT-5.6 Sol's 34.6, but it has the highest Agents' Last Exam score among the same batch of models, and requires significantly fewer output tokens to complete similar tasks.

What is the open-source license? The weights use Z.ai's custom GLM-5.3 License (no longer MIT like GLM-5.2), allowing commercial use, with ultra-large-scale MaaS scenarios requiring security review; the code portion is Apache 2.0.

What are the context and output limits? 1 million token context, maximum output of 128K tokens, text-only input. For image input, use GLM-5.3-Flash.

Can thinking mode be turned off? No. GLM-5.3's thinking mode is always on; you can only adjust the depth via reasoning_effort between low/high/max.

Does it support tool calls? Yes. Function calling, streaming output, and JSON structured output are all supported.

Why Use GLM-5.3 API with SeaWhale AI

Call directly without deployment — the 753B parameter model requires no self-built GPU cluster
OpenAI-compatible interface — connect with just two lines of changes to existing code
Side-by-side comparison with closed-source models — directly compare effectiveness and cost with Claude and GPT under the same account
Settlement in RMB — domestic models are priced in RMB, with clear accounting

API

API integration

Model IDUse this value as the model in inference requests

glm-5.3

API KeyBearer token used to authenticate inference requests

Manage API keys

Base URLOpenAI compatible · /chat/completions

https://api.haijingai.com/v2/

provider OptionalSelects a service channel; omit it and the system picks the default

"provider": { "channel": "direct" }

glm-5.3 usage examples

SeaWhale AI is compatible with the OpenAI API protocol, so you can call it with the OpenAI SDK or plain HTTP requests. Streaming is enabled by default.

cURLPythonNode.js

curl https://api.haijingai.com/v2/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "glm-5.3",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello!"}
    ],
    "provider": { "channel": "direct" },
    "stream": true
  }'
# provider is optional — remove this line to use the default channel

from openai import OpenAI

client = OpenAI(
    base_url="https://api.haijingai.com/v2",
    api_key="<API_KEY>",
)

stream = client.chat.completions.create(
    model="glm-5.3",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    stream=True,
    # Optional: pick a service channel; omit to use the default
    extra_body={"provider": {"channel": "direct"}},
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.haijingai.com/v2',
  apiKey: '<API_KEY>',
})

const stream = await client.chat.completions.create({
  model: 'glm-5.3',
  messages: [
    { role: 'system', content: 'You are a helpful assistant.' },
    { role: 'user', content: 'Hello!' },
  ],
  stream: true,
  // Optional: pick a service channel; omit to use the default
  // @ts-expect-error provider is a SeaWhale AI extension, not in the OpenAI SDK types
  provider: { channel: 'direct' },
})

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
}