zhipuText Chat

GLM 5.2 API: thinking-first model for research, coding, and structured answers

Try GLM 5.2 through AIReiter for thinking-heavy Chinese and English tasks, structured output, technical research, and multi-step reasoning.

InputAIReiter $1.40 per 1M tokensOutputAIReiter $4.40 per 1M tokensCache readAIReiter $0.26 per 1M tokens

Run with API

INPUT

FormJSONNode.jsPythonHTTP

OUTPUT

Example

PreviewJSON

A rate limit caps how many requests an API accepts from you in a given window. Once you exceed it, the server stops doing work for you and answers 429 Too Many Requests instead.

Handling 429

Read the Retry-After response header. When present it tells you exactly how long to wait, in seconds.
When it is absent, back off exponentially with jitter so retries from many clients do not line up.
Cap the number of retries, then surface the failure instead of looping forever.

async function withRetry(request, maxRetries = 4) {
  for (let attempt = 0; ; attempt++) {
    const response = await request();
    if (response.status !== 429 || attempt === maxRetries) return response;
    const retryAfter = Number(response.headers.get("retry-after"));
    const backoff = Number.isFinite(retryAfter) ? retryAfter * 1000 : 2 ** attempt * 500 + Math.random() * 250;
    await new Promise((resolve) => setTimeout(resolve, backoff));
  }
}

Treat the limit as a budget you plan around, not an error you retry your way out of: batch requests where you can, cache repeated reads, and spread bulk work over time.

{
  "model": "glm-5.2",
  "input": {
    "model": "glm-5.2",
    "max_tokens": 2048,
    "messages": [
      {
        "role": "user",
        "content": "Analyze this product decision and return a concise recommendation with tradeoffs."
      }
    ],
    "temperature": 0.7,
    "top_p": 1
  },
  "output": "A rate limit caps how many requests an API accepts from you in a given window. Once you exceed it, the server stops doing work for you and answers **429 Too Many Requests** instead.\n\n## Handling 429\n\n1. Read the `Retry-After` response header. When present it tells you exactly how long to wait, in seconds.\n2. When it is absent, back off exponentially with jitter so retries from many clients do not line up.\n3. Cap the number of retries, then surface the failure instead of looping forever.\n\n```js\nasync function withRetry(request, maxRetries = 4) {\n  for (let attempt = 0; ; attempt++) {\n    const response = await request();\n    if (response.status !== 429 || attempt === maxRetries) return response;\n    const retryAfter = Number(response.headers.get(\"retry-after\"));\n    const backoff = Number.isFinite(retryAfter) ? retryAfter * 1000 : 2 ** attempt * 500 + Math.random() * 250;\n    await new Promise((resolve) => setTimeout(resolve, backoff));\n  }\n}\n```\n\nTreat the limit as a budget you plan around, not an error you retry your way out of: batch requests where you can, cache repeated reads, and spread bulk work over time.",
  "metrics": {
    "input_tokens": 26,
    "output_tokens": 214,
    "generated_in_seconds": 4.1
  },
  "example": true
}

Generated in

4.1 seconds

Input tokens

Output tokens

214

Tokens per second

52.20 tokens / second

Time to first token

Model details

Use the same model key in Playground, API requests, and internal workflows.

Model ID: glm-5.2
Provider: zhipu
Protocol: Anthropic Messages
Context window: 1,000,000 tokens
Max output: -
Input tokens: 140 credits / 1M tokens
Output tokens: 440 credits / 1M tokens
Cache read: 26 credits / 1M tokens
Cache write: -

A reasoning model for structured decisions

GLM 5.2 is a good candidate when the answer needs a visible plan, a defensible recommendation, and stronger Chinese-English reasoning than a lightweight model.

Should you choose GLM 5.2?

Use it for research, decision support, technical analysis, and structured answers where the model may spend tokens thinking before producing visible text.

Choose it when

You need multi-step reasoning, Chinese-English analysis, structured recommendations, or technical research with explicit tradeoffs.

Use another model when

The job is a short batch extraction or latency-sensitive support response; thinking-heavy output can be overkill.

Public API protocol

Call POST https://aireiter.com/api/v1/messages with model "glm-5.2". Streaming is supported through the same Messages-compatible endpoint.

Token and cache usage

Usage may include thinking-heavy output. Always inspect input, output, and cache-read token fields instead of assuming visible text equals total billed output.

GLM 5.2 production workloads

Best for tasks where thinking budget improves the final answer.

Research synthesis

Merge notes, product facts, and competing claims into a recommendation with uncertainty called out.

Technical planning

Break a vague engineering problem into assumptions, risks, and next actions.

Structured analysis

Return comparisons, decision tables, and clearly labeled conclusions.

Bilingual knowledge work

Handle Chinese and English source material in the same reasoning flow.

How GLM 5.2 fits your model stack

Do not route every request to the newest model. Pick the cheapest model that still passes your quality bar, then reserve deeper models for failures or high-risk tasks.

For fast batches

Use Doubao or DeepSeek V4 Flash for fast batches; reserve GLM 5.2 for reasoning-heavy cases.

For deeper reasoning

Use GLM 5.2 when you want a deliberate answer; use DeepSeek V4 Pro for code-heavy deep reasoning.

For long context

Use MiniMax M3 or Kimi K2.7 Code if the prompt needs much more retained context.

For production rollout

Keep max_tokens reasonably high during evaluation so thinking-heavy responses are not cut off before visible text appears.

GLM 5.2 API questions

Questions developers usually check before moving a text model from playground testing to production API traffic.

/ 01

What model ID should I send for GLM 5.2?

Use "glm-5.2" in the API request body. The internal DB key is only used by AIReiter routing.

/ 02

Which endpoint should GLM 5.2 use?

Use POST https://aireiter.com/api/v1/messages for public API calls. Keep x-api-key / Authorization authentication consistent with your AIReiter API key setup.

/ 03

Does GLM 5.2 support streaming?

Yes. Send stream=true and read server-sent events until the message completes. Test non-streaming first when debugging authentication or model ID issues.

/ 04

How do I confirm token and cache billing for GLM 5.2?

Check the usage object returned by the API. Input, output, and cache-read token fields are the source of truth for settlement; a repeated prompt alone does not prove a cache hit.

/ 05

Should I always set max_tokens for GLM 5.2?

Do not set max_tokens too low. Some requests may spend output budget on reasoning before visible text, so leave enough room for the final answer.