Anthropic API Compatible

Call the model through an Anthropic-compatible Messages API.

POST/anthropic/v1/messages

Authorization

This endpoint supports the following two authentication methods. Choose one and add it to the request headers:

Method 1:x-api-keyauthentication, format:
x-api-key: $UNISOUND_API_KEY
Method 2:Authorization: Bearerauthentication, format:
Authorization: Bearer $UNISOUND_API_KEY

$UNISOUND_API_KEY Used to verify account information, can be viewed in Metered & Packs > API Key Management.

Request Bodyapplication/json

modelstringRequired

Model code. Available values:

Unisound Series: u2-flash, u2, u2-med, u2-radimed

DeepSeek Series: deepseek-v4-flash-0731, deepseek-v4-pro

Kimi Series: kimi-k2.5, kimi-k2.6, kimi-k3

GLM Series: glm-5.3, glm-5.2, glm-5.1, glm-5

MiniMax Series: MiniMax-M3, MiniMax-M2.5

Qwen (Tongyi Qianwen) Series: qwen3.8-flash, qwen3.7-plus, qwen3.7-max, qwen3.8-max, qwen3.6-plus, qwen3.6-flash, qwen3.6-35b-a3b

max_tokensintegerRequired

Maximum number of tokens to generate. max_tokens does not limit the thinking process; when deep thinking is enabled, thinking tokens are controlled separately by thinking.budget_tokens.

systemstring | arrayOptional

System prompt used to define the model role or behavior.

system is passed as a top-level parameter, and the messages array does not accept the system role.

Passing a string is equivalent to a single content block with type="text".

typestringRequired

Fixed to text.

textstringRequired

System prompt text.

messagesarrayRequired

Message array arranged in alternating user / assistant turns.

The messages array does not accept the system role.

rolestringRequired

Message role. Available values: user, assistant.

contentstring | arrayRequired

Message content. It can be a plain text string or a structured content array (including multimodal blocks such as text and images).

When content is a string, it is equivalent to a single content block with type="text".

Text informationobject

typestringRequired

Fixed to text.

textstringRequired

Text content.

Image informationobject

(requires a vision model)

typestringRequired

Fixed to image.

sourceobjectRequired

Source of the image data.

typestringRequired

Available values:

url: public image URL.
base64: Base64-encoded image data.

urlstringOptional

Public URL of the image. Required when type is url.

media_typestringOptional

MIME type of the image, such as image/jpeg. Required when type is base64.

datastringOptional

Base64-encoded image data. Required when type is base64.

Tool call informationobject

(assistant role, tool call instructions returned by the model)

typestringRequired

Fixed to tool_use.

idstringRequired

Unique identifier of the tool call, used to associate the result in a later tool_result.

namestringRequired

Name of the invoked tool.

inputobjectRequired

Input arguments of the tool call. The structure is determined by the corresponding input_schema in tools. The tool call content itself participates in the cache prefix.

Tool result informationobject

(user role, tool execution result sent back to the model)

typestringRequired

Fixed to tool_result.

tool_use_idstringRequired

Corresponds to the id in the tool_use information.

contentstringRequired

Content returned by the tool execution.

streambooleanOptionalDefault: false

Whether to enable streaming output. Defaults to false.

temperaturenumberOptional

Controls the diversity of generated text, with a value range of [0, 2). The larger the value, the more random the result.

This range differs from Anthropic's official [0.0, 1.0]. Confirm the value when migrating from Anthropic.

top_pnumberOptional

Nucleus sampling threshold. It is recommended to use either this or temperature, not both.

Value range: (0, 1.0].

The default value is determined by the model.

repetition_penaltynumberOptional

Penalty for consecutive sequence repetition. Gateway extension field, validated and forwarded upstream.

Value range: (0, 2.0].

Default value: 1.0 means no penalty.

presence_penaltynumberOptional

Controls content repetition. Gateway extension field, validated and forwarded upstream.

Value range: [-2.0, 2.0].

The default value is determined by the model.

frequency_penaltynumberOptional

Controls token frequency penalty. Gateway extension field.

Value range: [-2.0, 2.0].

The default value is determined by the model.

top_kintegerOptional

Size of the candidate sampling set during generation.

stop_sequencesarrayOptional

Stop sequences. Generation stops before a matched sequence, and the matched sequence itself is not included in the output.

When matched, the response stop_reason is still end_turn.

thinkingobjectOptional

Deep thinking configuration. When enabled, the model reasons before generating the reply to improve accuracy. The response includes content blocks of type thinking.

u2, kimi-k3, glm-5.3, and MiniMax-M2.5 enable thinking by default and do not support disabling it. MiniMax-M3 supports disabling thinking (thinking.type=disabled), unlike MiniMax-M2.5. u2-radimed disables thinking by default and does not support enabling it.

typestringRequired

Available values: enabled (thinking mode on), disabled (thinking mode off).

budget_tokensintegerOptional

Maximum number of tokens available for the thinking process. This does not overlap with max_tokens: this parameter limits thinking, while max_tokens limits the final reply. Takes effect when type is enabled.

toolsarrayOptional

Array of tool definitions for Function Call scenarios.

namestringRequired

Tool name.

descriptionstringOptional

Functional description of the tool.

input_schemaobjectRequired

JSON Schema definition of the tool input parameters.

tool_choiceobjectOptional

Tool selection strategy.

typestringRequired

Available values: auto (the model decides whether to call a tool, default), any (force the model to call any tool), none (disable tool calls), tool (force the model to call a specific tool).

namestringOptional

When type is tool, specifies the name of the tool to call.

reasoning_effortstringOptional

Controls the model's reasoning effort. Supported values and defaults vary by model.

Model	Default	Supported levels	Mapping of other levels
`deepseek-v4-pro`, `glm-5.2`, `glm-5.1`, `glm-5`	`high`	`high`, `max`	`low` and `medium` map to `high`; `xhigh` maps to `max`
`deepseek-v4-flash-0731`	`high`	`max`, `high`, `low`	`medium` maps to `high`; `xhigh` maps to `high`
`kimi-k3`, `glm-5.3`, `glm-5.3-flash`	`max`	`max`, `high`, `low`	—
`qwen3.8-max`, `qwen3.8-flash`	`xhigh`	`xhigh`, `medium`, `low`	`max` and `high` map to `xhigh`; `minimal` maps to `low`; `none` disables thinking
`u2-med`	—	`none`, `low`, `medium`, `high`	`minimal` maps to `low`; `xhigh` and `max` map to `high`

kimi-k3, glm-5.3, and glm-5.3-flash always enable thinking. thinking.type only supports enabled; passing disabled causes the API request to fail.

Values other than the supported options and mapped values above will cause an error.

output_configobjectOptional

Structured output configuration. When enabled, the model outputs a JSON string.

Strict structured output: applicable to DeepSeek Series and GLM Series models. The model strictly follows the provided JSON Schema for constrained output, ensuring field types and hierarchy are fully consistent.

Standard structured output: for all other models. Specific field constraints in the Schema are not enforced by default; the API converts it to standard JSON mode (valid JSON string only). Requires both: 1) explicitly passing output_config; 2) the word JSON (case-insensitive) in system or messages content.

formatobjectRequired

Output format configuration.

typestringRequired

Fixed to json_schema.

schemaobjectRequired

JSON Schema object. Must include type, properties, required, and additionalProperties (must be set to false).

Non-streaming response

idstring

Unique identifier of the message.

typestring

Fixed to message.

rolestring

Fixed to assistant.

modelstring

Model name used.

contentarray

Content array.

Text informationobject

typestring

Fixed to text.

textstring

Text reply generated by the model.

When output_config is enabled, this field is a JSON string that the client must parse.

Thinking informationobject

(returned when deep thinking is enabled)

typestring

Fixed to thinking.

thinkingstring

Model thinking process before generating the final reply.

signaturestring

Currently fixed to an empty string.

Tool call informationobject

(Function Call scenario)

typestring

Fixed to tool_use.

idstring

Unique identifier of the tool call, used to associate the result in a later tool_result.

namestring

Name of the invoked tool.

inputobject

Input arguments of the tool call.

stop_reasonstring

Stop reason. Available values:

end_turn (normal completion)
max_tokens (token limit reached)
tool_use (tool call)

usageobject

Token usage statistics.

In streaming mode, the usage field in the message_start event contains only input_tokens and output_tokens; all four fields are returned in the message_delta event.

input_tokensinteger

Number of input tokens.

output_tokensinteger

Number of output tokens.

cache_creation_input_tokensinteger

Number of input tokens consumed for cache creation.

cache_read_input_tokensinteger

Number of input tokens consumed for cache reads.

Streaming response

message_start

The first event of the stream, marking the start of the message.

typestring

Fixed to message_start.

messageobject

Initial message object. content is an empty array, and usage contains only input_tokens and output_tokens.

content_block_start

Sent when each content block starts, marking the index and type of the new content block.

typestring

Fixed to content_block_start.

indexinteger

Content block index, starting from 0, corresponding to the position in the message content array.

content_blockobject

Initial object of the content block. type can be text, thinking, or tool_use. For tool_use, input is an empty object in this event, and the full arguments are assembled from subsequent content_block_delta increments.

content_block_delta

Incremental update event for a content block. Multiple events may be sent for the same block.

typestring

Fixed to content_block_delta.

indexinteger

Index of the related content block.

deltaobject

Incremental object. type can be:

text_delta: text increment with the text field.
thinking_delta: thinking increment with the thinking field.
signature_delta: signature increment with the signature field (currently fixed to an empty string).
input_json_delta: tool call argument increment with the partial_json field.

content_block_stop

Event indicating the end of a content block.

typestring

Fixed to content_block_stop.

indexinteger

Index of the ended content block.

message_delta

Message-level update event sent after all content blocks end. It includes the stop reason and complete token usage statistics.

typestring

Fixed to message_delta.

deltaobject

Contains stop_reason and stop_sequence. See the non-streaming response table above for possible values.

usageobject

Complete token usage statistics, including input_tokens, output_tokens, cache_creation_input_tokens, and cache_read_input_tokens.

message_stop

The final event of the stream, marking the end of the message.

typestring

Fixed to message_stop.

Messages (Anthropic-compatible)

curl -X POST "https://maas-api.unisound.com/anthropic/v1/messages" \
  -H "Content-Type: application/json" \
  -H "x-api-key: $UNISOUND_API_KEY" \
  -d '{
    "model": "u2-flash",
    "max_tokens": 1024,
    "messages": [
        {
            "role": "user",
            "content": "你是谁？"
        }
    ],
    "thinking": {"type": "disabled"}
}'

200

{
  "id": "msg_f2517b5e-0f5e-4443-8f92-b9b53d8a60dc",
  "type": "message",
  "role": "assistant",
  "model": "u2-flash",
  "content": [
    {
      "type": "thinking",
      "thinking": "用户用中文问"你是谁？"，这是问我的身份。根据系统提示，我的名字是U2 Flash，由云知声（Unisound）构建。我应该用中文回答。",
      "signature": ""
    },
    {
      "type": "text",
      "text": "你好！我是U2 Flash，由云知声（Unisound）构建的AI助手。有什么我可以帮助你的吗？"
    }
  ],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {
    "input_tokens": 41,
    "output_tokens": 65,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0
  }
}