> ## Documentation Index
> Fetch the complete documentation index at: https://evolink.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# GLM All-Model API - Responses Reference
> Call GLM in Responses format for text conversations, streaming and function calling, with image understanding and web search available according to the model. See the parameters and model differences below.
**BaseURL**: The default BaseURL is `https://direct.evolink.ai`, which has better support for text models and long-lived connections. `https://api.evolink.ai` is the primary endpoint for multimodal services and serves as a fallback address for text models.
Call GLM in Responses format for text conversations, streaming and function calling, with image understanding and web search available according to the model. See the parameters and model differences below.
## Models and parameter differences
Choose a GLM model. All four support text on this endpoint. Optional capabilities vary by model.
| Model ID | Input | Reasoning notes |
| ---------------- | ------------ | -------------------------------------------------------------------------------------------------------------------- |
| `glm-5.3` | Text | Effective levels: `low` / `high` / `max`; compatible values are listed under reasoning. Thinking cannot be disabled. |
| `glm-5.3-flash` | Text, images | Same as `glm-5.3`; use `input_image` for images. |
| `glm-5.3-flashx` | Text, images | Same as `glm-5.3`; use `input_image` for images. |
| `glm-5.2` | Text | `none` may still produce reasoning tokens and does not guarantee that thinking is disabled. |
Responses uses nested `reasoning.effort`, rather than top-level `reasoning_effort` or thinking. Reasoning usage is included in output\_tokens. Simple tasks may return `reasoning_tokens=0`; this does not imply that thinking can be disabled.
Reasoning effort; `low` is recommended.
**Compatibility rules for `glm-5.3` / `glm-5.3-flash` / `glm-5.3-flashx`**
| Supplied value | Effective thinking level |
| ---------------------- | ------------------------------- |
| `low` / `high` / `max` | Unchanged |
| `xhigh` | `max` |
| `medium` | `high` |
| `minimal` / `none` | `low`; thinking remains enabled |
**`minimal` and `none` do not disable thinking on the 5.3 series.** Thinking tokens are billed as output. Unrecognized values are unchanged and have no compatibility mapping; use the listed values. These mappings do not apply to glm-5.2.
On this endpoint, `glm-5.2` may still produce reasoning tokens with none. This value does not guarantee that thinking is disabled.
## System prompts and multi-turn conversations
`instructions`: System instructions. `glm-5.3-flash` supports this field when input is a string. When input is a message array, put the system prompt in the first `role=system` message.
```json theme={null}
{
"model": "glm-5.3-flash",
"input": [
{
"role": "system",
"content": "Answer concisely in English."
},
{
"role": "user",
"content": "Remember the passphrase RED-583"
},
{
"role": "assistant",
"content": "Remembered"
},
{
"role": "user",
"content": "What is the passphrase? Reply with only the passphrase."
}
],
"reasoning": {
"effort": "low"
},
"max_output_tokens": 1024
}
```
Store the response for later reference. `glm-5.3-flash` and `glm-5.3-flashx` support multi-turn conversations using `store=true` with previous\_response\_id. `glm-5.2` does not support response-ID continuation; `store=true` does not enable it. Include the full history in input instead.
The top-level id of the previous response. `glm-5.3-flash` and `glm-5.3-flashx` support this with `store=true` and the same model. Pass the response id unchanged, not an item id from output. `glm-5.2` returns 400 for this field. For conversations across models, omit it and include the full history in input.
## Streaming responses
Enable SSE streaming. Read answer text from delta in response.output\_text.delta. The successful terminal event is response.completed. Also end the turn and handle `response.incomplete`, `response.failed` or error. Do not wait only for \[DONE] or connection closure.
| Event | Handling |
| ------------------------------------------------------------------------- | --------------------------------------------------------------------------------- |
| `response.created` / `response.in_progress` | The turn starts. |
| `response.output_text.delta` | Append delta to the answer text. |
| `response.reasoning_text.delta` / `response.reasoning_summary_text.delta` | Reasoning content; keep it separate from the answer and support both event forms. |
| `response.output_item.done` | Collect the complete output item, such as function\_call. |
| `response.completed` | Generation has ended; read `response.output` and response.usage. |
| `response.incomplete` / `response.failed` / `error` | Handle truncation or errors and end the turn. |
Stop reading after a terminal event. HTTP 200 only means that the stream was established; check the final event status. A tool-call turn may end with `response.completed` and still require your application to execute the function and send another request.
## Function calling
Choose the function example in the request menu. Responses function definitions use a flat structure:
```json theme={null}
{
"type": "function",
"name": "get_temperature",
"description": "Return the temperature for the specified city",
"parameters": {
"type": "object",
"properties": {
"city": {
"type": "string"
}
},
"required": [
"city"
],
"additionalProperties": false
}
}
```
1. Traverse `response.output` and collect all items with type=function\_call.
2. Parse and validate the arguments JSON string, then execute each function in your application.
3. Append the entire previous output to the history. Add a `function_call_output` for each call, using the original `call_id` and a string output.
4. Send the updated history as input in the next request. The function-result example shows this structure.
`parallel_tool_calls`: Whether multiple tools may be called in a turn. false does not guarantee a single function call. Clients should traverse and handle every function\_call.
## Images, search and JSON output
For `glm-5.3-flash` and `glm-5.3-flashx`, mix `input_text` and `input_image` in the user message content array. Pass a public image URL or Base64 Data URL in image\_url. Use text only for `glm-5.3` and glm-5.2.
Declare `tools: [{"type":"web_search"}]`. Search runs on the server and returns `web_search_call` items and answer text. Check the output items to see whether search was used. Searches may incur per-call fees in addition to token charges; see model pricing.
To continue after a search, append the entire previous `output`, including `web_search_call` and `message`, to `input`, then add your new question. Preserve original fields such as `id`, `status`, and `action`. The server has already executed the search, so do not create a `function_call_output` for `web_search_call`. See the `web_search_history` request example.
`text.format.type`: Output format: text for plain text or `json_object` for a JSON object. With `json_object`, explicitly request valid JSON in the prompt and parse and validate it in your client. Strict JSON Schema constraints are not provided; do not rely on `json_schema` or `strict=true` to enforce a structure.
## Responses and usage
Ordered output items. Extract text from content entries with `type=output_text` inside `type=message` items. Reasoning may precede the answer, and `function_call` turns may contain no answer text. Do not always read output\[0].
`output_text`: Optional aggregated answer text; it may be absent. General-purpose clients should traverse output.
`output_text` in message items; possibly `reasoning_text` in reasoning items. Reasoning may also be returned as summary\_text. Do not assume every reasoning item has content.
* `usage.input_tokens`: Total input tokens, including cached tokens. `usage.input_tokens_details.cached_tokens`: Cached subset of input tokens; do not add it to `input_tokens` again. Prefix caching is automatic and requires no explicit cache\_control. Use the returned value for the number of cache hits.
* `usage.output_tokens`: Total output tokens, including reasoning tokens. `usage.output_tokens_details.reasoning_tokens`: Subset of output tokens used for reasoning; do not count it again in output\_tokens. This detail may be absent or zero.
`status=incomplete` with `incomplete_details.reason=max_output_tokens` means the budget was exhausted. There may be reasoning without an answer; increase the output limit.
## OpenAPI
````yaml en/api-manual/language-series/glm/responses/responses-reference.json POST /v1/responses
openapi: 3.1.0
info:
title: GLM All-Model API - Responses Reference
description: >-
Call the Zhipu GLM series using the OpenAI-compatible Responses format.
Supports glm-5.3, glm-5.3-flash, glm-5.3-flashx and glm-5.2. Optional
capabilities vary by model.
license:
name: MIT
version: 1.0.0
servers:
- url: https://direct.evolink.ai
description: Production (recommended)
- url: https://api.evolink.ai
description: Alternative URL
security:
- bearerAuth: []
tags:
- name: Responses
description: GLM Responses API
paths:
/v1/responses:
post:
tags:
- Responses
summary: GLM Responses API (full reference)
description: >-
Call GLM in Responses format for text conversations, streaming and
function calling, with image understanding and web search available
according to the model. See the parameters and model differences below.
operationId: createGLMResponse
requestBody:
required: true
content:
application/json:
schema:
$ref: '#/components/schemas/ResponsesRequest'
examples:
basic:
summary: Basic text conversation
value:
model: glm-5.3-flash
input: Introduce yourself in one sentence.
max_output_tokens: 1024
reasoning:
effort: low
system_message:
summary: System prompt and message array
value:
model: glm-5.3-flash
input:
- role: system
content: Answer concisely in English.
- role: user
content: What is binary search?
max_output_tokens: 1024
reasoning:
effort: low
stream:
summary: SSE streaming output
value:
model: glm-5.3-flash
input: Introduce yourself in one sentence.
max_output_tokens: 1024
reasoning:
effort: low
stream: true
history:
summary: Multi-turn conversation with history
value:
model: glm-5.3-flash
input:
- role: user
content: Remember the passphrase RED-583
- role: assistant
content: Remembered
- role: user
content: What is the passphrase? Reply with only the passphrase.
max_output_tokens: 1024
reasoning:
effort: low
function:
summary: Client-side function calling
value:
model: glm-5.3-flash
input: Find the temperature in Beijing.
max_output_tokens: 1024
reasoning:
effort: low
tools:
- type: function
name: get_temperature
description: Return the temperature for the specified city
parameters:
type: object
properties:
city:
type: string
required:
- city
additionalProperties: false
tool_choice:
type: function
name: get_temperature
function_result:
summary: Return a function result
description: >-
Replace the example call_id with the actual value returned in
the previous turn and preserve the complete previous output.
value:
model: glm-5.3-flash
input:
- role: user
content: Find the temperature in Beijing.
- type: function_call
call_id: call_weather_demo
name: get_temperature
arguments: '{"city":"Beijing"}'
- type: function_call_output
call_id: call_weather_demo
output: '{"city":"Beijing","temperature":25}'
max_output_tokens: 1024
reasoning:
effort: low
tools:
- type: function
name: get_temperature
description: Return the temperature for the specified city
parameters:
type: object
properties:
city:
type: string
required:
- city
additionalProperties: false
tool_choice: none
web_search:
summary: Server-side web search
value:
model: glm-5.3-flash
input: >-
Use web search to find the title of the Python homepage.
Return only the title and source link.
max_output_tokens: 1024
reasoning:
effort: low
tools:
- type: web_search
tool_choice: required
web_search_history:
summary: Continue after a web search
description: >-
For an actual request, replace the sample web_search_call and
message with all original output items from the previous
response, then append your new question. Preserve fields such
as id, status, and action. Do not create a
function_call_output for web_search_call.
value:
model: glm-5.3-flash
input:
- role: user
content: >-
Use web search to find the title of the Python homepage.
Return only the title and source link.
- type: web_search_call
id: ws_search_demo
status: completed
action:
type: open_page
url: https://www.python.org
- type: message
id: msg_search_demo
status: completed
role: assistant
content:
- type: output_text
text: Welcome to Python.org — https://www.python.org
annotations: []
- role: user
content: >-
What homepage title did you find in the previous turn?
Return only the title without searching again.
max_output_tokens: 1024
reasoning:
effort: low
json_object:
summary: JSON object output
value:
model: glm-5.3-flash
input: >-
Return only a valid JSON object with a city field set to
Beijing. Do not use Markdown.
max_output_tokens: 1024
reasoning:
effort: low
text:
format:
type: json_object
image:
summary: Image understanding (Flash example)
description: The example is a Data URL for a red PNG image.
value:
model: glm-5.3-flash
max_output_tokens: 1024
reasoning:
effort: low
input:
- role: user
content:
- type: input_text
text: >-
The image is a solid color. Reply with only its
color.
- type: input_image
image_url: >-
data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAIAAAACACAIAAABMXPacAAABK0lEQVR4nO3RMQEAMAyAsLb+PW8y8hADHOybSEfraYDWAKwBWAOwBmANwBqANQBrANYArAFYA7AGYA3AGoA1AGsA1gCsAVgDsAZgDcAagDUAawDWAKwBWAOwBmANwBqANQBrANYArAFYA7AGYA3AGoA1AGsA1gCsAVgDsAZgDcAagDUAawDWAKwBWAOwBmANwBqANQBrANYArAFYA7AGYA3AGoA1AGsA1gCsAVgDsAZgDcAagDUAawDWAKwBWAOwBmANwBqANQBrANYArAFYA7AGYA3AGoA1AGsA1gCsAVgDsAZgDcAagDUAawDWAKwBWAOwBmANwBqANQBrANYArAFYA7AGYA3AGoA1AGsA1gCsAVgDsAZgDcAagDUAawDWAKwBWAOwBmANGOsDz+EB/5Uf+TQAAAAASUVORK5CYII=
store:
summary: Create a response for later reference (Flash)
value:
model: glm-5.3-flash
input: >-
Remember the passphrase BLUE-728. Reply only with
Remembered.
max_output_tokens: 1024
reasoning:
effort: low
store: true
previous_response:
summary: Reference the previous response (Flash)
description: >-
First run the example that creates a stored response, then set
previous_response_id to its top-level id. glm-5.2 does not
support this way of continuing a conversation.
value:
model: glm-5.3-flash
input: What was the passphrase? Reply with only the passphrase.
max_output_tokens: 1024
reasoning:
effort: low
previous_response_id: Response ID returned in the previous turn
flashx:
summary: Call GLM-5.3-FlashX
value:
model: glm-5.3-flashx
input: Introduce yourself in one sentence.
max_output_tokens: 1024
reasoning:
effort: low
vision_flashx:
summary: GLM-5.3-FlashX image input
value:
model: glm-5.3-flashx
max_output_tokens: 1024
reasoning:
effort: low
input:
- role: user
content:
- type: input_text
text: >-
The image is a solid color. Reply with only its
color.
- type: input_image
image_url: >-
data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAIAAAACACAIAAABMXPacAAABK0lEQVR4nO3RMQEAMAyAsLb+PW8y8hADHOybSEfraYDWAKwBWAOwBmANwBqANQBrANYArAFYA7AGYA3AGoA1AGsA1gCsAVgDsAZgDcAagDUAawDWAKwBWAOwBmANwBqANQBrANYArAFYA7AGYA3AGoA1AGsA1gCsAVgDsAZgDcAagDUAawDWAKwBWAOwBmANwBqANQBrANYArAFYA7AGYA3AGoA1AGsA1gCsAVgDsAZgDcAagDUAawDWAKwBWAOwBmANwBqANQBrANYArAFYA7AGYA3AGoA1AGsA1gCsAVgDsAZgDcAagDUAawDWAKwBWAOwBmANwBqANQBrANYArAFYA7AGYA3AGoA1AGsA1gCsAVgDsAZgDcAagDUAawDWAKwBWAOwBmANGOsDz+EB/5Uf+TQAAAAASUVORK5CYII=
responses:
'200':
description: >-
Generation completed or returned an incomplete result; check status.
Streaming returns text/event-stream.
content:
application/json:
schema:
$ref: '#/components/schemas/ResponsesResponse'
example:
id: response_demo
object: response
created_at: 1789971757
model: glm-5.3-flash
status: completed
output:
- type: message
id: message_demo
status: completed
role: assistant
content:
- type: output_text
text: >-
Hello, I am GLM. I can help with conversation, writing
and coding.
annotations: []
usage:
input_tokens: 17
output_tokens: 24
total_tokens: 41
input_tokens_details:
cached_tokens: 0
output_tokens_details:
reasoning_tokens: 0
error: null
text/event-stream:
schema:
type: string
example: >+
event: response.output_text.delta
data:
{"type":"response.output_text.delta","item_id":"message_demo","output_index":0,"content_index":0,"delta":"Hello"}
event: response.completed
data:
{"type":"response.completed","response":{"id":"response_demo","object":"response","created_at":1789971757,"model":"glm-5.3-flash","status":"completed","output":[{"type":"message","id":"message_demo","status":"completed","role":"assistant","content":[{"type":"output_text","text":"Hello","annotations":[]}]}],"usage":{"input_tokens":17,"output_tokens":3,"total_tokens":20},"error":null}}
'400':
description: >-
Invalid request parameters, such as missing input, malformed
reasoning settings or a model that does not support
previous_response_id.
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
'401':
description: The API key is invalid or expired.
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
'402':
description: Insufficient available credits.
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
'429':
description: Rate limit exceeded. Retry with backoff.
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
'500':
description: Server error.
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
'503':
description: The service is temporarily unavailable. Try again later.
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
components:
schemas:
ResponsesRequest:
type: object
properties:
model:
type: string
description: >-
Choose a GLM model. All four support text on this endpoint. Optional
capabilities vary by model.
| Model ID | Input | Reasoning notes |
| --- | --- | --- |
| `glm-5.3` | Text | Effective levels: low / high / max; compatible
values are listed under reasoning. Thinking cannot be disabled. |
| `glm-5.3-flash` | Text, images | Same as glm-5.3; use input_image
for images. |
| `glm-5.3-flashx` | Text, images | Same as glm-5.3; use input_image
for images. |
| `glm-5.2` | Text | none may still produce reasoning tokens and
does not guarantee that thinking is disabled. |
enum:
- glm-5.3
- glm-5.3-flash
- glm-5.3-flashx
- glm-5.2
default: glm-5.3-flash
example: glm-5.3-flash
input:
description: >-
Required. A text string or an array of Responses input items. Arrays
accept messages, returned model output items and
function_call_output. For multi-turn conversations, include the full
history in each request. Put the system prompt first as a
role=system message. Images use input_image, supported only by
glm-5.3-flash and glm-5.3-flashx. Do not use the Chat Completions
messages / image_url block format.
oneOf:
- type: string
- type: array
items:
$ref: '#/components/schemas/InputItem'
example: Introduce yourself in one sentence.
max_output_tokens:
type: integer
minimum: 1
description: >-
Maximum output tokens for this generation, including reasoning.
Start at 1024 and adjust for the task. A small budget may be
exhausted during reasoning, leaving only reasoning items and no
answer. Check status and incomplete_details. The parameter name is
max_output_tokens, not max_tokens.
example: 1024
stream:
type: boolean
default: false
description: >-
Enable SSE streaming. Read answer text from delta in
response.output_text.delta. The successful terminal event is
response.completed. Also end the turn and handle
response.incomplete, response.failed or error. Do not wait only for
[DONE] or connection closure.
reasoning:
type: object
properties:
effort:
type: string
description: >-
Reasoning effort; low is recommended.
**Compatibility rules for glm-5.3 / glm-5.3-flash /
glm-5.3-flashx**
| Supplied value | Effective thinking level |
| --- | --- |
| `low` / `high` / `max` | Unchanged |
| `xhigh` | `max` |
| `medium` | `high` |
| `minimal` / `none` | low; thinking remains enabled |
**minimal and none do not disable thinking on the 5.3 series.**
Thinking tokens are billed as output. Unrecognized values are
unchanged and have no compatibility mapping; use the listed
values. These mappings do not apply to glm-5.2.
On this endpoint, glm-5.2 may still produce reasoning tokens
with none. This value does not guarantee that thinking is
disabled.
enum:
- max
- xhigh
- high
- medium
- low
- minimal
- none
example: low
description: >-
Responses uses nested reasoning.effort, rather than top-level
reasoning_effort or thinking. Reasoning usage is included in
output_tokens. Simple tasks may return reasoning_tokens=0; this does
not imply that thinking can be disabled.
instructions:
type: string
description: >-
System instructions. glm-5.3-flash supports this field when input is
a string. When input is a message array, put the system prompt in
the first role=system message.
tools:
type: array
items:
$ref: '#/components/schemas/Tool'
description: >-
Supports client-side function tools and server-side web_search.
Declare functions with flat name / description / parameters fields,
not a nested Chat Completions function object. Your application
executes function_call and returns the result. web_search runs on
the server; actual searches may incur per-call charges in addition
to tokens. See model pricing.
tool_choice:
description: >-
auto lets the model choose; none disables tools; required requires a
tool call. To select a function, use
{"type":"function","name":"get_temperature"}. Forced selection
behavior is not guaranteed to be identical for every model and tool
combination.
oneOf:
- type: string
enum:
- auto
- none
- required
- type: object
properties:
type:
type: string
const: function
name:
type: string
required:
- type
- name
example: auto
parallel_tool_calls:
type: boolean
description: >-
Whether multiple tools may be called in a turn. false does not
guarantee a single function call. Clients should traverse and handle
every function_call.
text:
type: object
properties:
format:
type: object
properties:
type:
type: string
description: >-
Output format: text for plain text or json_object for a JSON
object. With json_object, explicitly request valid JSON in
the prompt and parse and validate it in your client. Strict
JSON Schema constraints are not provided; do not rely on
json_schema or strict=true to enforce a structure.
enum:
- text
- json_object
example: json_object
required:
- type
description: >-
Output format. Examples use json_object. HTTP 200 does not guarantee
JSON Schema compliance.
store:
type: boolean
description: >-
Store the response for later reference. glm-5.3-flash and
glm-5.3-flashx support multi-turn conversations using store=true
with previous_response_id. glm-5.2 does not support response-ID
continuation; store=true does not enable it. Include the full
history in input instead.
previous_response_id:
type: string
description: >-
The top-level id of the previous response. glm-5.3-flash and
glm-5.3-flashx support this with store=true and the same model. Pass
the response id unchanged, not an item id from output. glm-5.2
returns 400 for this field. For conversations across models, omit it
and include the full history in input.
example: Response ID returned in the previous turn
metadata:
type: object
additionalProperties:
type: string
description: >-
Custom string key-value metadata, available in response metadata. Do
not include keys or sensitive information.
example:
conversation: demo
temperature:
type: number
description: >-
Sampling parameter. The effective range and behavior depend on the
model. It does not guarantee deterministic output and can be omitted
for reasoning tasks.
top_p:
type: number
description: >-
Sampling parameter. The effective range and behavior depend on the
model. It can usually be omitted.
required:
- model
- input
ResponsesResponse:
type: object
properties:
id:
type: string
description: ID of this response. Pass it unchanged as previous_response_id.
example: response_demo
object:
type: string
const: response
created_at:
type: integer
description: Creation time in Unix seconds.
model:
type: string
example: glm-5.3-flash
status:
type: string
description: >-
completed means generation for this turn has ended, possibly with
tool calls only. incomplete means the output is incomplete. Check
both output and error.
enum:
- completed
- incomplete
- failed
- in_progress
- queued
output:
type: array
items:
$ref: '#/components/schemas/OutputItem'
description: >-
Ordered output items. Extract text from content entries with
type=output_text inside type=message items. Reasoning may precede
the answer, and function_call turns may contain no answer text. Do
not always read output[0].
output_text:
type: string
description: >-
Optional aggregated answer text; it may be absent. General-purpose
clients should traverse output.
usage:
$ref: '#/components/schemas/Usage'
error:
type:
- object
- 'null'
description: Response error, normally null on success.
additionalProperties: true
incomplete_details:
type: object
properties:
reason:
type: string
description: Details when output is truncated, for example max_output_tokens.
metadata:
type:
- object
- 'null'
additionalProperties:
type: string
ErrorResponse:
type: object
properties:
error:
type: object
properties:
message:
type: string
type:
type: string
param:
type:
- string
- 'null'
code:
type:
- string
- integer
- 'null'
required:
- error
InputItem:
description: >-
A message, function result or an item copied unchanged from the previous
output. For tool results, use the returned call_id and preserve the
original fields of previous output items.
oneOf:
- $ref: '#/components/schemas/InputMessage'
- $ref: '#/components/schemas/FunctionCallOutput'
- type: object
properties:
type:
type: string
description: Type of the returned output item.
enum:
- function_call
- reasoning
- web_search_call
id:
type: string
call_id:
type: string
name:
type: string
arguments:
type: string
description: Arguments encoded as a JSON string.
status:
type: string
action:
type: object
additionalProperties: true
description: Search or page-access action for web_search_call.
required:
- type
Tool:
oneOf:
- type: object
properties:
type:
type: string
const: function
name:
type: string
description:
type: string
parameters:
type: object
description: JSON Schema for function parameters.
strict:
type: boolean
description: >-
Function-argument constraint option. Clients must still parse
and validate arguments before executing the function.
required:
- type
- name
- parameters
- type: object
properties:
type:
type: string
const: web_search
required:
- type
OutputItem:
type: object
properties:
type:
type: string
description: >-
Common types are message, reasoning, function_call and
web_search_call.
enum:
- message
- reasoning
- function_call
- web_search_call
id:
type: string
status:
type: string
role:
type: string
content:
type: array
items:
type: object
properties:
type:
type: string
text:
type: string
annotations:
type: array
items:
type: object
description: >-
output_text in message items; possibly reasoning_text in reasoning
items.
summary:
type: array
items:
type: object
properties:
type:
type: string
text:
type: string
description: >-
Reasoning may also be returned as summary_text. Do not assume every
reasoning item has content.
call_id:
type: string
description: Function call identifier used to return its result.
name:
type: string
description: Function name.
arguments:
type: string
description: >-
Function arguments as a JSON string. Parse and validate them before
execution.
action:
type: object
additionalProperties: true
description: Search or page-access action for web_search_call.
required:
- type
Usage:
type: object
properties:
input_tokens:
type: integer
description: Total input tokens, including cached tokens.
output_tokens:
type: integer
description: Total output tokens, including reasoning tokens.
total_tokens:
type: integer
description: Sum of input and output tokens.
input_tokens_details:
type: object
properties:
cached_tokens:
type: integer
description: >-
Cached subset of input tokens; do not add it to input_tokens
again. Prefix caching is automatic and requires no explicit
cache_control. Use the returned value for the number of cache
hits.
output_tokens_details:
type: object
properties:
reasoning_tokens:
type: integer
description: >-
Subset of output tokens used for reasoning; do not count it
again in output_tokens. This detail may be absent or zero.
InputMessage:
type: object
properties:
role:
type: string
enum:
- system
- user
- assistant
content:
description: >-
A text string or an array of input content blocks. When sending back
an existing assistant output, its output_text blocks can be
preserved unchanged.
oneOf:
- type: string
- type: array
items:
oneOf:
- $ref: '#/components/schemas/InputText'
- $ref: '#/components/schemas/InputImage'
- $ref: '#/components/schemas/OutputText'
required:
- role
- content
FunctionCallOutput:
type: object
properties:
type:
type: string
const: function_call_output
call_id:
type: string
description: The call_id of the original function_call.
output:
type: string
description: Function result, usually a JSON-encoded string.
required:
- type
- call_id
- output
InputText:
type: object
properties:
type:
type: string
const: input_text
text:
type: string
required:
- type
- text
InputImage:
type: object
properties:
type:
type: string
const: input_image
image_url:
type: string
description: >-
A public image URL or Base64 Data URL, such as
data:image/png;base64,... for PNG. Only glm-5.3-flash /
glm-5.3-flashx support images; use text only for glm-5.3 and
glm-5.2.
required:
- type
- image_url
OutputText:
type: object
properties:
type:
type: string
const: output_text
text:
type: string
annotations:
type: array
items:
type: object
required:
- type
- text
securitySchemes:
bearerAuth:
type: http
scheme: bearer
description: Send Bearer YOUR_API_KEY in the Authorization header.
````