GLM 5.3

GLM 5.3 (z-ai/glm-5.3) is a large-scale reasoning model from Z.ai built for complex software engineering and long-horizon agent tasks. Better than GLM-5.2 at coding and the quality-per-token balance: reasoning is always on (Low/High/Max), with a 1M-token context. Text-only.

Parameters and capabilities

Parameter	Value
model	z-ai/glm-5.3 (or the short glm-5.3)
Context	1M tokens
Reasoning	always on — cannot be disabled
reasoning_effort	low, high, max
Image input	no — an image_url part returns 400
File input	no — a file part returns 400
Cache read	29 ₽ per 1M tokens
Cache write	not published in the catalog

Reasoning is innate to GLM 5.3: reasoning_effort picks the depth (low/high/max), while off and medium are not supported by the model — they never reach the upstream. If you omit the parameter entirely, the default depth from the model card applies — for GLM 5.3 it is the maximum.

Examples

Code review

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://api.mixen.ai/v1",
    api_key="mxn-...",
)

resp = client.chat.completions.create(
    model="z-ai/glm-5.3",
    messages=[{
        "role": "user",
        "content": "Here is a service class: <code>. Find race conditions and resource leaks, propose fixes",
    }],
)
print(resp.choices[0].message.content)

cURL

curl https://api.mixen.ai/v1/chat/completions \
  -H "Authorization: Bearer $MIXEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.3",
    "messages": [{
      "role": "user",
      "content": "Here is a service class: <code>. Find race conditions and resource leaks, propose fixes"
    }]
  }'

Reasoning depth per task

low saves tokens on simple fixes; max turns on the full depth for architectural work:

curl https://api.mixen.ai/v1/chat/completions \
  -H "Authorization: Bearer $MIXEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.3",
    "reasoning_effort": "low",
    "messages": [{"role": "user", "content": "Write a SQL migration: add an audit table partitioned by month"}]
  }'

Context caching in /v1/chat/completions kicks in automatically on a repeating prefix — there are no control parameters. For an agent session this means: keep the system prompt and tool descriptions stable from request to request.

Price

Per 1M tokens: input 156 ₽ ($1.75), output 490.2 ₽ ($5.5), cache read 29 ₽ ($0.325); a separate cache-write price is not published by the catalog. Current prices — catalog.

Strengths and limits

• The model’s profile is complex software engineering and long-horizon agent tasks; at coding and the quality-per-token balance it is better than GLM-5.2.
• 1M-token context — whole repositories and long session histories fit in.
• Reasoning depth is adjustable: low, high, max.
• Reasoning cannot be turned off: reasoning tokens are always part of the bill — budget for them.
• Reasoning eats into max_tokens: set a generous ceiling for long answers, or you will get finish_reason: "length".
• Text only: images and files are not accepted as input — a request with them returns 400.

All models — in the catalog. The general text-model workflow — chat guide.