> ## 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 - Messages Reference
> - Call GLM series models through the Anthropic Messages protocol, choosing the specific model with the `model` parameter
- The request and response structures match the Anthropic API
- **System prompt**: passed through the top-level `system` field
- **Thinking mode**: thinking is on by default across the series and is returned in `content[type=thinking]` blocks; only `glm-5.2` can turn it off with `thinking.type=disabled`
- **Streaming**: SSE event stream
- **Tool calling**: compatible with the Anthropic `tool_use` / `tool_result` flow
- **Image input**: genuinely supported only by `glm-5.3-flash`; see the `messages` field for details
**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.
**Whether thinking can be turned off varies by model**: only `glm-5.2` supports `thinking.type: "disabled"`; `glm-5.3` and `glm-5.3-flash` always think and return an error if `disabled` is sent. When migrating from `glm-5.2`, remove the hard-coded `disabled` before switching models.
**Image input is supported only by `glm-5.3-flash`, and sending images to the other models does not raise an error.** When image content blocks go to `glm-5.3` or `glm-5.2`, the request returns 200 as usual, but the model cannot read the image and answers from the text alone — a plausible-looking reply that has nothing to do with the image, and one that is not consistent across requests. This kind of silent failure is hard to track down in production, so choose `glm-5.3-flash` when you need image understanding.
## OpenAPI
````yaml en/api-manual/language-series/glm/messages/messages-reference.json POST /v1/messages
openapi: 3.1.0
info:
title: GLM All-Model API - Messages Reference
description: >-
Full API reference for calling Zhipu GLM text models through the Anthropic
Messages protocol.
**Models covered**: `glm-5.3`, `glm-5.3-flash`, `glm-5.2` (selected with the
`model` parameter)
**Compatibility notes**:
- Path: `/v1/messages` (the standard Anthropic path)
- The request and response structures match the Anthropic Messages API
- Supported fields: `model` `messages` (required) `system` `max_tokens`
`temperature` `top_p` `top_k` `stop_sequences` `stream` `thinking` `tools`
`tool_choice` `metadata`
**Shared capabilities**:
- 1M token context window, up to 131,072 tokens of output
- Thinking mode: **thinking is on by default** across the series; it is
returned in `content[type=thinking]` blocks and counts towards output tokens
- Prompt caching: an implicit prefix cache that repeated requests with the
same prefix hit automatically (reported in `cache_read_input_tokens`), with
no need to set `cache_control`; `cache_creation_input_tokens` is always 0
**Differences between models** (whether thinking can be turned off, whether
images can be read) are described in the `model`, `thinking` and `messages`
fields below.
license:
name: MIT
version: 1.0.0
servers:
- url: https://direct.evolink.ai
description: Production (Recommended, better support for text models)
- url: https://api.evolink.ai
description: Alternative URL
security:
- bearerAuth: []
tags:
- name: Messages
description: Anthropic Messages protocol endpoints
paths:
/v1/messages:
post:
tags:
- Messages
summary: GLM Messages (All Models, Anthropic-Compatible)
description: >-
- Call GLM series models through the Anthropic Messages protocol,
choosing the specific model with the `model` parameter
- The request and response structures match the Anthropic API
- **System prompt**: passed through the top-level `system` field
- **Thinking mode**: thinking is on by default across the series and is
returned in `content[type=thinking]` blocks; only `glm-5.2` can turn it
off with `thinking.type=disabled`
- **Streaming**: SSE event stream
- **Tool calling**: compatible with the Anthropic `tool_use` /
`tool_result` flow
- **Image input**: genuinely supported only by `glm-5.3-flash`; see the
`messages` field for details
operationId: createMessageGLM
requestBody:
required: true
content:
application/json:
schema:
$ref: '#/components/schemas/CreateMessageRequest'
examples:
simple:
summary: Minimal runnable request
value:
model: glm-5.3
max_tokens: 1024
messages:
- role: user
content: Hello, world
system_prompt:
summary: With a system prompt
value:
model: glm-5.3
max_tokens: 2048
system: You are a seasoned technical editor.
messages:
- role: user
content: Introduce GLM-5.3 in three sentences.
stop_sequences:
summary: Custom stop sequences
value:
model: glm-5.3
max_tokens: 50
thinking:
type: disabled
stop_sequences:
- '3'
messages:
- role: user
content: 'Output exactly: 1 2 3 4 5 6'
tool_use:
summary: Tool calling (Anthropic tool_use style)
value:
model: glm-5.3
max_tokens: 2048
messages:
- role: user
content: Check the weather in Tokyo and tell me
tools:
- name: get_weather
description: Query the current weather of a specified city
input_schema:
type: object
properties:
city:
type: string
description: 'City name, for example: Tokyo'
required:
- city
tool_choice:
type: auto
streaming:
summary: Streaming output (SSE)
value:
model: glm-5.3
max_tokens: 1024
stream: true
messages:
- role: user
content: Write a short poem about spring
disable_thinking_glm52_only:
summary: Disable thinking mode (glm-5.2 only)
value:
model: glm-5.2
max_tokens: 512
thinking:
type: disabled
messages:
- role: user
content: 'In one sentence: what is the capital of Japan?'
description: >-
Only `glm-5.2` can turn thinking off. Sending
`thinking.type=disabled` to `glm-5.3` or `glm-5.3-flash`
returns an error.
vision_flash:
summary: Image input (glm-5.3-flash only)
description: >-
`glm-5.3-flash` supports vision natively. Images are passed in
through `image` content blocks, and `source` accepts both
`base64` and `url`.
value:
model: glm-5.3-flash
max_tokens: 1024
messages:
- role: user
content:
- type: text
text: What is in this image?
- type: image
source:
type: base64
media_type: image/png
data:
responses:
'200':
description: Message object
content:
application/json:
schema:
$ref: '#/components/schemas/MessageResponse'
examples:
with_thinking:
summary: Includes a thinking content block by default
value:
id: msg_0842a705-9d0b-4eaa-b12d-09a4106326c5
type: message
role: assistant
model: glm-5.3
content:
- type: thinking
thinking: >-
The user asked to greet them with one word, so
answering "Hi" will do.
signature: ''
- type: text
text: Hi.
stop_reason: end_turn
usage:
input_tokens: 18
output_tokens: 101
cache_creation_input_tokens: 0
cache_read_input_tokens: 0
prompt_tokens_details:
cached_tokens: 0
tool_use:
summary: Triggers a tool call (stop_reason=tool_use)
value:
id: msg_067e85db-53df-43a1-bd38-09c53375f2f0
type: message
role: assistant
model: glm-5.3
content:
- type: tool_use
id: toolu_36b8a98e284c426799f08612
name: get_weather
input:
city: Tokyo
stop_reason: tool_use
usage:
input_tokens: 161
output_tokens: 11
cache_creation_input_tokens: 0
cache_read_input_tokens: 0
prompt_tokens_details:
cached_tokens: 0
'400':
description: Invalid request parameters
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
type: error
request_id: req_xxx
error:
type: invalid_request_error
message: Invalid request
'401':
description: Authentication error
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
type: error
error:
type: authentication_error
message: Authentication error
'402':
description: Insufficient quota
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
type: error
error:
type: billing_error
message: Insufficient quota
'403':
description: Permission error
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
type: error
error:
type: permission_error
message: Permission denied
'404':
description: Model or resource not found
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
type: error
error:
type: not_found_error
message: Model not found
'429':
description: Rate limited
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
type: error
error:
type: rate_limit_error
message: Rate limited
'500':
description: Internal server error
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
'502':
description: Upstream service error
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
'503':
description: Service temporarily unavailable
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
components:
schemas:
CreateMessageRequest:
type: object
required:
- model
- messages
properties:
model:
type: string
description: >
Model to call:
| Model ID | Positioning | Thinking can be turned off | Image input
|
|---|---|---|---|
| `glm-5.3` | Flagship model, with across-the-board gains on complex
software engineering and agent tasks; 1M context | Cannot be turned
off | Not supported |
| `glm-5.3-flash` | Lightweight multimodal model, extremely low cost
with native vision; 1M context | Cannot be turned off | Supported |
| `glm-5.2` | Previous-generation flagship, complex reasoning and
very long context; 1M context | Can be turned off
(`thinking.type=disabled`) | Not supported |
enum:
- glm-5.3
- glm-5.3-flash
- glm-5.2
example: glm-5.3
default: glm-5.3
max_tokens:
type: integer
description: >-
Upper limit on the length of the generated content (in tokens)
**Note**:
- The GLM series supports up to **131,072 tokens** (128K) of output;
setting at least `1024` is recommended
- Tokens produced by thinking also count towards this limit
- When the limit is reached the content is truncated and the
response carries `stop_reason=max_tokens`
minimum: 1
maximum: 131072
example: 1024
messages:
type: array
description: >-
The list of conversation messages, alternating between user and
assistant turns
**Note**:
- At least one message is required
- The last message is normally `role=user`
- Multi-turn context is supported and the model draws on the history
**Image input**: only `glm-5.3-flash` supports it, through a
`{"type":"image","source":{...}}` block inside the `content` array.
Sending image content blocks to `glm-5.3` or `glm-5.2` **does not
return an error, but the model cannot read the image**. The request
returns 200 as usual, and the model answers from the text alone,
producing a plausible-looking reply that has nothing to do with the
image — and the result is not consistent across requests.
This kind of silent failure is hard to track down in production, so
choose `glm-5.3-flash` when you need image understanding.
items:
$ref: '#/components/schemas/InputMessage'
minItems: 1
system:
description: >-
System prompt, used to set the AI's role and behavior
**Notes**:
- Supports a string or an array of content blocks
- Passed via the top-level `system` field (do not place it inside
messages)
- The model follows the system constraints
- **An overly long system may be truncated**: For long context,
place it in `messages` rather than piling everything into `system`
oneOf:
- type: string
example: You are a helpful assistant.
- type: array
description: >-
System prompt in content-block array format. text blocks may
carry cache_control
items:
type: object
required:
- type
- text
properties:
type:
type: string
enum:
- text
text:
type: string
cache_control:
$ref: '#/components/schemas/CacheControl'
temperature:
type: number
description: >-
Sampling temperature
**Notes**:
- Higher values make output more varied, lower values more
deterministic
- Recommended range `[0, 1]`
minimum: 0
maximum: 1
example: 1
top_p:
type: number
description: >-
Nucleus sampling threshold
**Notes**:
- Range `[0, 1]`
- It is recommended not to adjust temperature and top_p at the same
time
minimum: 0
maximum: 1
example: 0.9
top_k:
type: integer
description: >-
Sample only from the K highest-probability tokens (an
Anthropic-specific parameter)
**Notes**:
- Smaller values make output more deterministic, larger values make
candidates more diverse
minimum: 0
example: 10
stop_sequences:
type: array
description: >-
Custom stop sequences: generation stops when it hits any of these
strings
**Notes**:
- Hitting one truncates output, and content before the hit is
returned normally
- **Note**: When a stop sequence is hit, the GLM series returns
`stop_reason` as `end_turn` (rather than the Anthropic-standard
`stop_sequence`), and the response does not include a
`stop_sequence` field. If your client relies on
`stop_reason=="stop_sequence"` to detect a hit, special handling is
required
items:
type: string
example:
- |+
stream:
type: boolean
description: >-
Whether to return via SSE streaming
- `true`: Server-Sent Events streaming (standard Anthropic event
sequence: message_start / content_block_start / content_block_delta
/ message_delta / message_stop)
- `false`: Returns the complete response all at once (default)
default: false
example: false
thinking:
type: object
description: >-
Controls deep thinking
**Note**:
- Every model in the GLM series is a reasoning model, and **thinking
is on by default when this field is omitted**
- When it is on, the response `content` array contains a
`type="thinking"` reasoning block (billed as output tokens;
`signature` may be an empty string)
- **Only the binary `type` switch takes effect**: thinking budget or
level parameters such as `budget_tokens` and `effort` have no effect
(they are ignored)
**Whether it can be turned off depends on the model**:
- `glm-5.2`: sending `{"type":"disabled"}` turns thinking off and
cuts output tokens noticeably
- `glm-5.3` / `glm-5.3-flash`: **always think, cannot be turned
off**. Sending `disabled` returns an error
**The consequence: the `glm-5.3` series cannot reduce thinking
overhead on this endpoint.** It cannot be turned off (`disabled`
errors),
and it cannot be turned down either (`budget_tokens` and `effort`
both have no effect, and the top-level `reasoning_effort` is an
OpenAI-protocol field that this endpoint ignores).
Thinking content is billed as output tokens, so on this endpoint
that cost is unavoidable.
**To control thinking cost, switch to the [Chat Completions
API](../chat-completions/chat-completions-reference)** —
there, `reasoning_effort` has three levels that genuinely take
effect: `low` / `high` / `max`. `glm-5.2` is not subject to this
limit; it can turn thinking off directly on this endpoint.
**Migrating from `glm-5.2`**: if your code hard-codes
`thinking.type=disabled`, you must remove that field before
switching to `glm-5.3`, otherwise the request fails outright.
And if you relied on turning thinking off to control cost, this
endpoint offers no equivalent — plan to move to the Chat Completions
API as part of the switch.
required:
- type
properties:
type:
type: string
enum:
- enabled
- disabled
description: |-
- Not passing the thinking field: thinking is enabled by default
- `disabled`: turn off thinking and answer directly
- `enabled`: the Anthropic-standard explicit enable value
tools:
type: array
description: >-
The list of tool definitions
**Notes**:
- Follows the Anthropic tool definition spec
- `input_schema` uses a JSON Schema object
- The model returns standard `tool_use` blocks with
`stop_reason=tool_use`
items:
$ref: '#/components/schemas/Tool'
tool_choice:
type: object
description: Tool selection strategy
required:
- type
properties:
type:
type: string
enum:
- auto
- none
description: |-
- `auto`: the model decides on its own whether to call a tool
- `none`: tool calls are disallowed
metadata:
type: object
description: Request metadata
properties:
user_id:
type: string
description: >-
A unique identifier representing the end user, which can be used
for per-user monitoring and abuse detection (using a hashed ID
is recommended)
MessageResponse:
type: object
description: Anthropic-style message response
properties:
id:
type: string
description: 'The message''s unique ID (format: `msg_`)'
type:
type: string
enum:
- message
description: Response object type
role:
type: string
enum:
- assistant
model:
type: string
description: Model actually used
example: glm-5.3
content:
type: array
description: >-
The list of response content blocks
**Possible block types**:
- `thinking`: the reasoning process (when thinking is enabled, which
is the default)
- `text`: the final answer text
- `tool_use`: a tool call initiated by the model
items:
$ref: '#/components/schemas/OutputContentBlock'
stop_reason:
type: string
description: >-
Stop reason
- `end_turn`: natural completion (also returned when stop_sequences
is hit)
- `max_tokens`: reached the max_tokens limit
- `tool_use`: the model triggered a tool call
enum:
- end_turn
- max_tokens
- tool_use
usage:
$ref: '#/components/schemas/AnthropicUsage'
ErrorResponse:
type: object
properties:
type:
type: string
enum:
- error
error:
type: object
properties:
type:
type: string
description: >-
Error type (e.g. invalid_request_error / authentication_error /
billing_error, etc.)
message:
type: string
description: Error description
request_id:
type: string
description: Request tracing ID
InputMessage:
type: object
required:
- role
- content
properties:
role:
type: string
enum:
- user
- assistant
description: >-
The role of the message sender, alternating between user / assistant
- `user`: user message (also used to pass back `tool_result` blocks)
- `assistant`: the assistant's historical replies (may contain
`text` / `thinking` / `tool_use` blocks)
**`system` is not accepted**: pass system prompts via the top-level
`system` field.
content:
description: >-
Message content
**Note**:
- For plain text, pass a string directly
- For structured content, pass an array of content blocks (`text` /
`image` / `tool_use` / `tool_result` / `thinking`)
- `image` content blocks are **supported only by `glm-5.3-flash`**;
sending them to any other model does not return an error, but the
model cannot read the image
oneOf:
- type: string
- type: array
items:
$ref: '#/components/schemas/ContentBlock'
CacheControl:
type: object
description: >-
Prompt cache marker
**Note**: the GLM series uses an implicit cache (built automatically
from identical prefixes, hits shown in `cache_read_input_tokens`, needs
warming up) and does not rely on this explicit marker; `cache_control`
can be sent normally but may be ignored, and it does not affect the
implicit cache.
required:
- type
properties:
type:
type: string
enum:
- ephemeral
description: Cache marker type
Tool:
type: object
required:
- name
- input_schema
properties:
name:
type: string
description: |-
Tool name
**Notes**:
- Only `a-zA-Z0-9_-` are allowed
- Maximum 64 characters
description:
type: string
description: >-
A description of the tool's capability, for the model to decide when
to call it
input_schema:
type: object
description: |-
The JSON Schema object for the tool's input parameters
**Notes**:
- `type` must be `object`
- Should declare `properties` and `required`
cache_control:
$ref: '#/components/schemas/CacheControl'
OutputContentBlock:
type: object
description: A content block in the response
properties:
type:
type: string
enum:
- text
- thinking
- tool_use
text:
type: string
description: Text when type=`text`
thinking:
type: string
description: The reasoning-process text when type=`thinking`
signature:
type: string
description: Signature when type=`thinking` (may be an empty string)
id:
type: string
description: Tool call ID when type=`tool_use`
name:
type: string
description: Tool name when type=`tool_use`
input:
type: object
description: >-
The JSON input parameters generated by the model when
type=`tool_use`
AnthropicUsage:
type: object
description: Token usage statistics (Anthropic specification)
properties:
input_tokens:
type: integer
description: Number of input tokens (the portion not served from cache)
example: 18
output_tokens:
type: integer
description: Number of output tokens (including thinking)
example: 101
cache_creation_input_tokens:
type: integer
description: >-
Number of tokens written to the cache. The GLM series uses implicit
caching with no explicit cache-creation step, so **this field is
always 0**.
example: 0
cache_read_input_tokens:
type: integer
description: >-
Number of input tokens served from the implicit prefix cache.
Repeated requests with the same prefix hit it automatically, with no
need to set `cache_control`; the cached portion is billed at the
cache rate, which is markedly lower than the uncached input rate.
The first request returns 0.
example: 0
prompt_tokens_details:
type: object
description: >-
Input token breakdown (cache-hit fields, also returned by the GLM
series)
properties:
cached_tokens:
type: integer
description: Number of input tokens served from cache
example: 0
ContentBlock:
type: object
description: >-
Message content block
**Supported types**:
- `text`: text content
- `image`: image (**`glm-5.3-flash` only**; other models do not return
an error but cannot read the image)
- `tool_use`: replay of the previous assistant tool call
- `tool_result`: tool execution result
- `thinking`: replay of the previous assistant thinking content
required:
- type
properties:
type:
type: string
enum:
- text
- image
- tool_use
- tool_result
- thinking
text:
type: string
description: Text content when type=`text`
source:
type: object
description: >-
Image source (required when type=`image`, **supported only by
`glm-5.3-flash`**)
Both inline `base64` and `url` are supported.
required:
- type
properties:
type:
type: string
enum:
- base64
- url
description: >-
How the image is supplied
- `base64`: inline image data
- `url`: a public image URL, downloaded by the upstream service
itself
media_type:
type: string
description: Image MIME type (required when `type=base64`)
example: image/png
data:
type: string
description: >-
Base64-encoded image data (required when `type=base64`, without
the `data:` prefix)
example: iVBORw0KGgoAAAANSUhEUgAA...
url:
type: string
description: Public image URL (required when `type=url`)
example: https://example.com/photo.jpg
id:
type: string
description: Tool call ID (required for tool_use)
name:
type: string
description: Tool name (required for tool_use)
input:
type: object
description: Tool input parameters (for tool_use, a JSON object)
tool_use_id:
type: string
description: >-
The corresponding tool call ID (required for tool_result, filled
back from tool_use.id)
content:
description: >-
Tool execution result (tool_result), a string or an array of content
blocks
oneOf:
- type: string
- type: array
items:
type: object
thinking:
type: string
description: >-
The passed-back assistant thinking-process content (used when
type=`thinking`)
signature:
type: string
description: >-
The signature of the returned thinking content; it must be passed
back unchanged when continuing a multi-turn conversation (may be an
empty string)
cache_control:
$ref: '#/components/schemas/CacheControl'
securitySchemes:
bearerAuth:
type: http
scheme: bearer
description: >-
##All interfaces require authentication using a Bearer Token##
**Get an API Key**:
Visit the [API Key management page](https://evolink.ai/dashboard/keys)
to obtain your API Key
**Add it to the request header when using**:
```
Authorization: Bearer YOUR_API_KEY
```
**Note**: EvoLink uses Bearer Token authentication uniformly for
`/v1/messages`.
````