Chat Completions

The GLM 5 Chat Completions endpoint follows an OpenAI-compatible request and response shape for text chat. Existing OpenAI-compatible clients can usually keep the same SDK and request body, then change the base URL, API key, and model ID.

https://glm5.app/api/v1

Basic Request

curl https://glm5.app/api/v1/chat/completions \
  -H "Authorization: Bearer $GLM5_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.3-flash",
    "messages": [
      {"role": "system", "content": "Answer in concise Markdown."},
      {"role": "user", "content": "Explain idempotency in two bullets."}
    ],
    "max_completion_tokens": 600
  }'

Request Body

The request JSON body must be at most 4 MB. GLM 5 does not impose one global token context or output window across all models, and does not add its own ceiling on top of a model's native limits. max_completion_tokens is passed straight through to the selected model; omit it and the model runs to its own default stopping point. If a value is genuinely too large for the model you selected, the request fails the same way it would calling that model directly — see Errors for the shape of that response. Context and output limits vary by model; oversized requests return context_length_exceeded.

Safe Retries And Request Status

For network retries, send one stable Idempotency-Key with the request. GLM5 creates at most one inference task for each key and API key pair. The request body may be retried in any JSON key order, but reusing the key for different content returns 409 idempotency_key_reused.

Idempotency-Key: codex:turn_01JQ9KPC3H5YQ2T8

If the original request is still running, or has already reached a terminal state, GLM5 returns 409 with request_id, request_status, and status_url. It also exposes the same URL in X-GLM5-Request-Status. Poll that endpoint with the same API key instead of submitting the request again:

curl https://glm5.app/api/v1/requests/chatcmpl-01... \
  -H "Authorization: Bearer $GLM5_API_KEY"

The status response includes the terminal state, token usage, charged credits, and a safe error code when applicable. It does not retain or replay completion text, so a client reconnects to inspect task state and then decides whether a new request is needed. This avoids silently storing prompts or responses just to support retries.

For an in-flight request, streaming or non-streaming, send an explicit cancellation with the same API key:

curl -X POST https://glm5.app/api/v1/requests/chatcmpl-01.../cancel \
  -H "Authorization: Bearer $GLM5_API_KEY"

The endpoint returns 202 with cancel_requested, then the status endpoint settles as cancelled. Cancelled reservations are released; error remains reserved for provider, validation, and timeout failures.

Supported fields

GLM 5 accepts the documented Chat Completions fields below. Only the documented subset is guaranteed to affect model behavior.

Additional SDK fields

Some SDKs may send fields that are not listed in this reference. GLM 5 may ignore unknown fields for compatibility, but applications should not depend on undocumented behavior.

Notable current limitations include:

• response_format JSON mode and JSON Schema enforcement are not available.
• frequency_penalty, presence_penalty, repetition_penalty, and logit_bias are not applied.
• logprobs and top_logprobs are not returned.
• reasoning and reasoning_effort are not exposed as public controls.
• modalities, audio, and image_config are not supported by this text endpoint.
• stream_options, service_tier, prediction, and parallel_tool_calls are not currently applied.

For strict integrations, only depend on fields documented on this page.

Messages

messages is an ordered array. GLM 5 is stateless: previous turns are only remembered when your application sends them again.

Only text parts are used. Image, audio, video, and file parts are ignored by this text endpoint.

[
  {
    "role": "system",
    "content": "You are a concise technical assistant."
  },
  {
    "role": "user",
    "content": [
      {
        "type": "text",
        "text": "Summarize this deployment checklist."
      }
    ]
  }
]

When trimming history, keep assistant tool_calls and their matching tool messages together. A tool message without a preceding matching tool_call_id returns 400 invalid_request_error.

Non-Streaming Response

Non-streaming calls return a chat.completion object:

{
  "id": "chatcmpl-01...",
  "object": "chat.completion",
  "created": 1785292800,
  "model": "glm-5.3-flash",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "- Idempotency makes retries safe..."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 42,
    "completion_tokens": 31,
    "total_tokens": 73
  }
}

Response Fields

GLM 5 does not currently return fields such as usage.cost, usage.prompt_tokens_details.cached_tokens, service_tier, or system_fingerprint.

Model limits

GLM 5 does not return synthetic context-budget headers, does not publish a per-model context/output limit table, and does not enforce an application ceiling below what the selected model natively supports — capability limits come from the model itself. Use the final usage.prompt_tokens value for monitoring and handle any context or output-limit error returned for the selected model. A streaming (stream: true) response may run for up to 90 seconds before it is cut off with stream_timeout; see Errors.

Finish Reasons

For incremental output, see Streaming. For tool-call loops, see Function calling.