Create a Message

POST https://zenmux.ai/api/anthropic/v1/messages

ZenMux supports the Anthropic API. See the API call examples for how to use it.

This article is based on the official “Create a Message” documentation. Parameter descriptions and nested structures remain consistent with the official docs.

Request headers

x-api-key string

Anthropic API Key, used for authentication.

Example:

x-api-key: sk-ant-xxxx

anthropic-version string

Anthropic API version (not the model version).

Currently, only "2023-06-01" is supported.

anthropic-version: 2023-06-01

content-type string

Request body format. Currently, only JSON is supported:

content-type: application/json

anthropic-beta string

Used to enable one or more Beta features. "code-execution-2025-08-25" is not supported at the moment, meaning the code_execution tool cannot be used.

• For multiple beta versions, you can:
  • Separate them with commas: anthropic-beta: files-api-2025-04-14,another-beta
  • Or repeat the header multiple times.

Request Body

The request body is JSON. Parameters are as follows.

max_tokens number

The maximum number of tokens to generate, including both the normal response and (if enabled) the thinking content from extended thinking.

• Meaning: the model will generate at most this many tokens; it may end naturally earlier, but will not exceed this limit.
• The maximum supported max_tokens varies by model. See each model’s documentation for details.
• Value: >= 1

messages array<Message>

Conversation history and the current user input.

• During training, the model assumes user / assistant alternate. You provide prior turns in messages, and the model generates the next assistant message.
• Multi-turn conversations are supported; consecutive messages with the same role are merged internally.
• If the last entry in messages is assistant, the reply will continue directly after its content, which can be used for “prefix-constrained” answers.
• Up to 100000 messages per request.

All array elements share the same structure:

Message = {
  role: "user" | "assistant",
  content: string | ContentBlock[]
}

role: "user" | "assistant"

content string | ContentBlock[]

• If you provide a string directly, it is equivalent to a single text block with type: "text":

  • {"role":"user","content":"Hello, Claude"} is equivalent to
    {"role":"user","content":[{"type":"text","text":"Hello, Claude"}]}

• Using an array lets you mix multiple content block types, such as text, images, PDFs, tool results, etc.

  Common content block types (some are only available in specific scenarios):

TextBlockParam

{
  type: "text",
  text: string,
  cache_control?: CacheControlEphemeral,
  citations?: TextCitationParam[]
}

• text string
  Text content.

• type string
  Must be "text".

• cache_control CacheControlEphemeral
  Creates a Prompt Cache breakpoint on this block (for Anthropic context-caching billing / reuse):

  CacheControlEphemeral = {
    type: "ephemeral",
    ttl: "5m" | "1h",
  };

  • ttl: cache TTL, "5m" / "1h", default is 5 minutes.

• citations TextCitationParam[]
  Used to attribute the source of quoted text (typical scenario: when you pass a PDF / text document / content document as a document block input and want to mark “which page / which section / which search result this part of the answer comes from”).
  TextCitationParam is one of the following (depending on the cited content type):

  • char_location: cite plain text or content documents by character range

    {
      type: "char_location",
      cited_text: string,
      document_index: number,
      document_title: string,
      start_char_index: number,
      end_char_index: number
    }

    Field details:

    • type string
      Must be "char_location", meaning the citation is located by a character index range.

    • cited_text string
      The cited original text snippet (for human display).
      Typically extracted from the text between start_char_index and end_char_index.

    • document_index number
      The index of the cited document in the current request, 0-based.
      If you provide multiple document blocks (or other citeable documents) in this request, this indicates which one.

    • document_title string
      The document’s title/name, usually derived from the filename or a title you provided upstream, used for UI display such as “From: xxx”.

    • start_char_index number
      The starting character index (0-based, inclusive) of the cited snippet in the document’s full text.

    • end_char_index number
      The ending character index (0-based, typically the right endpoint of a half-open interval), so the cited range is [start_char_index, end_char_index).

  • page_location: cite PDFs by page range

    {
      type: "page_location",
      cited_text: string,
      document_index: number,
      document_title: string,
      start_page_number: number,
      end_page_number: number
    }

    Field details:

    • type string
      Must be "page_location", meaning the location is described by a page number range.

    • cited_text string
      The cited PDF text snippet (readable text parsed from the PDF by the system).

    • document_index number
      The index of the cited PDF in the current request, 0-based.

    • document_title string
      The PDF title or filename, for display.

    • start_page_number number
      The starting page number of the cited content, 1-based, inclusive.
      For example, 5 means “starting from page 5”.

    • end_page_number number
      The ending page number, usually treated as the right endpoint of a half-open interval:

      • If start_page_number = 5 and end_page_number = 6, it can be interpreted as “page 5”.
      • If the difference is greater than 1, it indicates a multi-page citation range.

  • content_block_location: cite content documents by content-block index range

    {
      type: "content_block_location",
      cited_text: string,
      document_index: number,
      document_title: string,
      start_block_index: number,
      end_block_index: number
    }

    Used to cite documents provided as “multiple content blocks” (e.g., a document with source.type = "content" that contains multiple text/image blocks).

    Field details:

    • type string
      Must be "content_block_location".

    • cited_text string
      The cited original text snippet (from the corresponding content block).

    • document_index number
      The index of the cited document in the current request, 0-based.

    • document_title string
      The document’s title/name.

    • start_block_index number
      The starting block index (0-based) within the document’s internal content array.
      Indicates “starting from which content block”.

    • end_block_index number
      The ending block index within the content array.
      In practice, this is usually interpreted as the other endpoint of the range:

      • If start_block_index === end_block_index, it typically indicates citing a single block.
      • If they differ, it indicates a citation range spanning multiple blocks.

  • web_search_result_location: cite Web search results

    {
      type: "web_search_result_location",
      cited_text: string,
      url: string,
      title: string,
      encrypted_index: string
    }

    Used when Anthropic’s Web Search tool (server tool) is enabled and Claude cites content from a webpage.

    Field details:

    • type string
      Must be "web_search_result_location", meaning this citation comes from Web search results.

    • cited_text string
      A snippet of cited webpage text (usually truncated for display) and does not count toward token usage.

    • url string
      The cited webpage URL, which the frontend can render as a clickable link.

    • title string
      The cited webpage title (e.g., HTML <title>), for UI display such as “Source: xxx”.

    • encrypted_index string
      An encrypted index identifier for this search result. It must be returned to Anthropic verbatim for follow-up turns to continue citing or inspecting the same result.
      You typically don’t need to show it to end users, but you must preserve it for multi-turn conversations / debugging.

  • search_result_location: cite custom retrieval results (SearchResultBlock)

    {
      type: "search_result_location",
      cited_text: string,
      source: string | null,
      title: string | null,
      search_result_index: number,
      start_block_index: number,
      end_block_index: number
    }

    When you provide your own search / RAG results to Claude via a type: "search_result" content block and enable citations, Claude will use this type when citing those results in its answer.

    Field details:

    • type string
      Must be "search_result_location", meaning the citation comes from a SearchResultBlock you provided.

    • cited_text string
      The exact cited text snippet, sourced from the text in a search_result content block.

    • source string | null
      The source identifier of the search result:

      • Usually a URL (e.g., a knowledge base document link);
      • Or a custom string ID you define;
      • May be null if you did not provide it in the original search_result.

    • title string | null
      The search result title, corresponding to search_result.title;
      If no title is available, it is null.

    • search_result_index number
      The 0-based index of the type: "search_result" block being cited within the current message.content.
      Whether you place these results in a user message or they are returned by a tool, they are numbered by appearance order.

    • start_block_index number
      The 0-based starting block index of the cited content within that search_result’s content array.

    • end_block_index number
      The ending block index within that content array.

      • If it equals start_block_index, it typically means a single content block is cited.
      • Otherwise, it indicates a citation range spanning multiple blocks.

ImageBlockParam

{
  type: "image",
  source: Base64ImageSource | URLImageSource,
  cache_control?: CacheControlEphemeral
}

• type string
  Must be "image".
• source Base64ImageSource | URLImageSource :
  • Base64ImageSource

    {
      type: "base64",
      media_type: "image/jpeg" | "image/png" | "image/gif" | "image/webp",
      data: string // base64-encoded
    }

  • URLImageSource

    {
      type: "url",
      url: string
    }

• cache_control CacheControlEphemeral : same as above; you can create a cache breakpoint for images.

DocumentBlockParam

{
  type: "document",
  source: Base64PDFSource | PlainTextSource | ContentBlockSource | URLPDFSource,
  cache_control?: CacheControlEphemeral,
  citations?: TextCitationParam[],
  context?: string,
  title?: string
}

• type string
  Must be "document".

• source Base64PDFSource | PlainTextSource | ContentBlockSource | URLPDFSource :

  • Base64PDFSource: base64 PDF

  {
    type: "base64",
    media_type: "application/pdf",
    data: string
  }

  • PlainTextSource: a full plain-text document

    {
      type: "text",
      media_type: "text/plain",
      data: string
    }

  • ContentBlockSource: provide a set of ContentBlocks as the document content (multimodal)

    {
      type: "content",
      content: string | ContentBlockSourceContent[]
    }

  • URLPDFSource: reference a remote PDF

    {
      type: "url",
      url: string
    }

• cache_control CacheControlEphemeral : same as above; you can create a cache breakpoint for documents.

• citations TextCitationParam[] : same as above; used to attribute the source of quoted text (typical scenario: provide PDFs / text documents / content documents as document blocks and mark “which page / which paragraph / which search result this part of the answer comes from”).

• context string : document context.

• title string : document title.

ToolResultBlockParam

{
  type: "tool_result",
  tool_use_id: string,          // Corresponds to the id in the previous tool_use block
  content?: string | (TextBlockParam | ImageBlockParam | SearchResultBlockParam | DocumentBlockParam)[]
  cache_control?: CacheControlEphemeral,
  is_error?: boolean
}

• type string
  Must be "tool_result".
• tool_use_id string , binds to the specific tool invocation.
• is_error boolean .
• content: either a simple string or an array of multimodal blocks (text / images / documents / search results) .
• cache_control CacheControlEphemeral : same as above.

ToolUseBlockParam

{
  type: "tool_use",
  id: string,                     // Unique tool invocation ID
  name: string,                   // Tool name defined in tools
  input: Record<string, unknown>, // JSON matching the tool's input_schema
  cache_control?: CacheControlEphemeral
}

Field details:

• type string
  Must be "tool_use".

• id string
  The unique identifier for this tool invocation, used to match the subsequent tool_result.

• name string
  The tool name to call. It must exactly match a tool.name declared in the request’s tools array.

• input object (Record<string, unknown>)

• cache_control CacheControlEphemeral : same as above, sets caching behavior (Prompt Cache breakpoint) for this tool invocation block.

ServerToolUseBlockParam

Indicates that Claude decides to call a server-side tool (hosted by Anthropic, not a client tool you implement), for example:

{
  type: "server_tool_use",
  id: string,
  name: string,
  input: Record<string, unknown>,
  cache_control?: CacheControlEphemeral
}

Field details:

• type string
  Must be "server_tool_use", indicating this is a server-side tool call request.

• id string
  The unique ID for this server tool invocation, like "srvtoolu_...".
  Subsequent result blocks (e.g., web_search_tool_result) will reference this ID via tool_use_id.

• name string
  The server tool name to call, for example:

  • "web_search": Web Search tool

• input object
  Parameters passed to the server tool. The structure is defined by the specific tool.

WebSearchToolResultBlockParam

When using the Web Search tool, Claude will return one or more web_search_tool_result content blocks within the same assistant message, representing “the web_search results or error for this call”.

{
  type: "web_search_tool_result",
  tool_use_id: string,
  content: `array | object`,
  cache_control?: CacheControlEphemeral
}

Field details:

• type string
  Must be "web_search_tool_result", indicating a Web Search tool result.

• tool_use_id string
  References the corresponding prior server_tool_use.id, used to associate the “search request” and the “search results”.

• cache_control CacheControlEphemeral : same as above.

• content array | object
  The execution result of Web Search:

  • On success: an array of web_search_result objects;
  • On failure: a web_search_tool_result_error object (see the error structure below).

web_search_result object (on success)

When content is an array, each element is a web_search_result object:

Field details:

• type string
  Must be "web_search_result".

• url string
  The URL of the result page, typically consistent with the url in citations.

• title string
  The page title, used for frontend display of the citation source.

• encrypted_content string
  An encrypted string of the page’s main content.
  In multi-turn conversations, if you want Claude to continue citing this result accurately, you need to return this field verbatim as part of the conversation (e.g., via a later web_search_result_location citation). This field is opaque and not parseable to you.

• page_age string
  An approximate last-update or crawl time, such as "April 30, 2025", mainly for user display of “data freshness”.

Error result structure: web_search_tool_result_error

If the Web Search tool itself errors (e.g., exceeds max uses, invalid request, etc.), the content field of web_search_tool_result will be an error object:

{
  "type": "web_search_tool_result",
  "tool_use_id": "servertoolu_a93jad",
  "content": {
    "type": "web_search_tool_result_error",
    "error_code": "max_uses_exceeded",
  },
}

Error object fields:

• type string
  Must be "web_search_tool_result_error".

• error_code string
  Error type code. Common values include:

  • "too_many_requests": the search tool hit a rate limit;
  • "invalid_input": invalid search parameters (e.g., invalid domain filters);
  • "max_uses_exceeded": exceeded the max_uses limit configured for this request;
  • "query_too_long": the generated search query is too long;
  • "unavailable": an internal error or temporary unavailability of the search service.

Even when an error occurs, the HTTP status code is still 200. The error is only reflected in the content of web_search_tool_result. You should decide how to degrade gracefully or notify users based on error_code.

ThinkingBlockParam

{
  type: "thinking",
  thinking: string,
  signature: string,
}

Field details:

• type string
  Must be "thinking", indicating an Extended Thinking reasoning block.

• thinking string
  Claude’s human-readable reasoning content, typically a multi-line step-by-step analysis.

• signature string
  An encrypted signature of the full thinking content, used in subsequent turns to verify that these reasoning blocks were generated by Claude and have not been tampered with.

  • This is an opaque field; you do not need to and should not parse it.
  • When you send back a previous assistant message that includes thinking to the API, you should include the full thinking + signature verbatim.

In streaming mode:

• thinking text is emitted incrementally via thinking_delta in content_block_delta events;
• signature is appended once via a signature_delta event before the block ends. You must concatenate all thinking_delta.thinking pieces, and then combine with the final signature as a complete thinking block.\

SearchResultBlockParam

Used to provide your own search / RAG results as structured content to Claude, making it easier for the model to cite them in its response and automatically generate search_result_location citations.

Typical scenario: you retrieve from a vector DB / document store on the backend, then place results into messages[*].content as search_result blocks.

{
  type: "search_result",
  source?: string,
  title?: string,
  content: Array<TextBlockParam | ImageBlockParam | DocumentBlockParam>,
  cache_control?: CacheControlEphemeral,
  citations?: {
    enabled: boolean
  }
}

Field details:

• type string
  Must be "search_result", indicating a search/retrieval result content block.

• source string
  A source identifier for the search result:

  • Usually a URL (e.g., knowledge base doc link, internal doc viewer link);
  • Or a custom string ID (e.g., document primary key);
  • If inconvenient to provide, you can omit it or set it to null.
    When Claude generates search_result_location citations, it will return this field verbatim so you can display “From: xxx” in the UI.

• title string
  The title of the search result:

  • Such as “API Reference: Authentication” or “Employee Handbook · Leave Policy”;
  • If no suitable title is available, it can be null.
    In citations, it is used directly as the citation title for UI rendering.

• content array
  The actual content snippets of the search result, consisting of one or more content blocks, typically text-heavy but can include images / documents, etc.

• cache_control CacheControlEphemeral : same as above.

• citations object

  Whether to enable automatic citation attribution based on this search result. Typical usage:

  citations: {
    enabled: boolean;
  }

  • enabled boolean
    • true: allow Claude to generate search_result_location citations for this search_result in the response;
    • false: do not generate citations for this block (the model can still read and use it).

RedactedThinkingBlockParam

RedactedThinkingBlockParam corresponds to a type: "redacted_thinking" content block and is part of the Extended Thinking system. Unlike a normal thinking block, its reasoning content is encrypted/redacted and not presented in plain text, mainly for safety/compliance while still allowing the model to reference its prior reasoning across turns.

You will typically only see it in model outputs and return it verbatim in subsequent requests; you rarely need to construct it yourself.

{
  type: "redacted_thinking",
  data: string
}

Field details:

• type string
  Must be "redacted_thinking", indicating a redacted thinking block.

  • Compared with type: "thinking":
    • thinking: returns readable natural-language reasoning text + signature;
    • redacted_thinking: returns unreadable encrypted data and does not include readable reasoning content.

• data string
  The encrypted/redacted thinking data string, usually a long Base64/ciphertext-like blob.

  • You cannot and do not need to parse this data.
  • Key point: to carry this thinking context forward into later turns, include the entire redacted_thinking block verbatim in a new request as part of the prior assistant messages.

model string

The model ID used for this call.

• This is a ZenMux-defined model name, for example:
  • "anthropic/claude-sonnet-4.5"

Note: This differs from the Anthropic API-style <model> string.

output_config object

Configuration options for model output, such as output format.

• format object a schema used to specify Claude’s response output format.
  • schema Record<string, unknown> the JSON schema for the format
  • type string must be "json_schema"

stop_sequences string[]

Custom stop sequences.

• When the generated text contains any stop sequence:
  • Generation stops immediately;
  • stop_reason = "stop_sequence" in the response;
  • stop_sequence is set to the matched string.
• If not set, the model stops at natural completion with end_turn.

Common usage:

• Use "END" as an end-of-answer marker;
• Combine with “multi-part output” protocols.

stream boolean

Whether to return results as a stream via SSE (Server-Sent Events).

• false (default): return a complete message object in one response.
• true: incrementally stream output via multiple events (see “Streaming Response” below).

system string | TextBlockParam[]

A System Prompt that sets global instructions and roles for the conversation. This is equivalent to Claude’s “overall rules” and takes effect before all messages.

Type formats

• You can provide a string directly (most common):

  "system": "You are a helpful assistant."

• Or you can provide an array of TextBlockParam, with the same structure as message:

  "system": [
    {
      "type": "text",
      "text": "You are a helpful assistant that answers in Chinese.",
      "cache_control": { "type": "ephemeral", "ttl": "1h" }
    },
    {
      "type": "text",
      "text": "The current date is 2025-01-15."
    }
  ]

Note: The Messages API does not have a role: "system" message. All system-level instructions are passed via the top-level system field.

temperature number

Sampling temperature, controlling output randomness.

• Default: 1.0
• Range: 0.0 ~ 1.0
  • Closer to 0: more deterministic, more “exam-like”; suitable for multiple-choice questions and rigorous reasoning.
  • Closer to 1: more diverse and creative; suitable for brainstorming and creative writing.
• Even 0.0 is not absolutely deterministic.

thinking object

Extended Thinking (explicit reasoning process) configuration.

thinking?:
  | { type: "enabled";  budget_tokens: number }
  | { type: "disabled" }

• type: "enabled":
• budget_tokens number:
  • Token budget allocated for internal reasoning;
  • Must be >= 1024 and < max_tokens;
  • Larger budgets typically improve reasoning quality for complex problems, but increase latency/cost.
  • The response will include content blocks with type: "thinking".
• type: "disabled": disables extended thinking (default behavior).

tool_choice object

Controls how Claude uses tools declared in tools.

tool_choice?:
  | { type: "auto";  disable_parallel_tool_use?: boolean }
  | { type: "any";   disable_parallel_tool_use?: boolean }
  | { type: "tool";  name: string; disable_parallel_tool_use?: boolean }
  | { type: "none" }

• "auto" (recommended default)
  • Claude decides whether to use tools and which tools to use;
  • disable_parallel_tool_use?: boolean:
    • false (default): allow multiple parallel tool_use blocks in a single response;
    • true: call at most 1 tool.
• "any"
  • Means “any tool may be used”; similar to "auto" but typically more strongly encourages tool usage;
  • disable_parallel_tool_use has the same meaning as above.
• "tool"
  • Forces the use of a specific tool:

    { type: "tool", name: "get_weather" }

  • If disable_parallel_tool_use is true, only this tool is called once.
• "none"
  • Disables tool usage; generate text/multimodal output only.

tools array<ToolUnion>

Declares the list of tools Claude can use in this request.

Officially, tools are divided into:

• Client tools: implemented by you in your application (similar to “function calling”)
• Server tools: hosted by Anthropic, such as Web Search, Bash, Text Editor, etc.

1. Custom (Client) Tool

The most basic JSON Schema tool definition:

{
  type?: "custom",              // Optional
  name: string,                 // Tool name (<= 128 chars)
  description?: string,         // Strongly recommended; the more detailed the better
  input_schema: {
    type: "object",
    properties?: { [key: string]: any },
    required?: string[]
  },
  cache_control?: CacheControlEphemeral
}

• name: Claude uses this name to call your tool in a tool_use block;
• description: clearly explain the tool purpose, parameter meanings, and usage constraints in natural language; this helps the model decide whether to call it and how to fill parameters correctly;
• input_schema: JSON Schema for the tool input;
• cache_control: define cache breakpoints for the tool.

Example tool_use block generated by Claude:

{
  "type": "tool_use",
  "id": "toolu_01D7FLrfh4G...",
  "name": "get_stock_price",
  "input": { "ticker": "^GSPC" }
}

After you run the tool, return the result in a tool_result block in the next user message.

2. Built-in Server Tools (selected)

The Messages API docs list several built-in tool types, typical examples include:

• Bash tool: type: "bash_20250124", name: "bash"

• Text editor: type: "text_editor_2025xxxx", name: "str_replace_editor" / "str_replace_based_edit_tool" etc.

  • Some versions include a max_characters field to cap the number of characters returned for display.

• Web Search tool:
  type: "web_search_20250305", name: "web_search"

  Configuration options include:

  {
    name: "web_search",
    type: "web_search_20250305",
    allowed_domains?: string[],
    blocked_domains?: string[],
    max_uses?: number,
    user_location?: {
      type: "approximate",
      city?: string,
      country?: string, // ISO 3166-1 alpha-2
      region?: string,
      timezone?: string // IANA time zone ID
    },
    cache_control?: CacheControlEphemeral
  }

For detailed semantics, invocation patterns, and billing details of each built-in tool, refer to Anthropic’s separate “Server tools” documentation.

top_k number

During sampling, choose only from the top K tokens with the highest probabilities.

• Used to truncate low-probability “long tail” tokens;
• Recommended only for advanced tuning; in most cases, temperature alone is sufficient;
• >= 0.

top_p number

Nucleus sampling parameter.

• Accumulate probability mass in descending order until the total reaches top_p, and sample only from that set;
• Range: 0.0 ~ 1.0;
• Typically used as an alternative to temperature; not recommended to adjust both significantly at the same time.

Unsupported fields

Field name	Type	Supported	Description
metadata	object	❌ Not supported	Business metadata for the request
service_tier	string	❌ Not supported	Service tier

Response

Non-streaming: returns a “complete message object”

When calling POST /v1/messages with stream: false (or omitted), Anthropic returns a complete Message object in one response. The following describes the field structure layer by layer.

{
  "id": "msg_013Zva2CMHLNnXjNJJKqJ2EF",
  "type": "message",
  "role": "assistant",
  "model": "claude-sonnet-4-5-20250929",
  "content": [ ... ],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": { ... }
}

id string

The unique ID of this message.

type "message"

Object type. Fixed to "message" for the Messages API.

role "assistant"

The author role of the message generated by Claude. Always "assistant".

model string

The model name that actually processed this request (same as or equivalent to the model in the request).

content array<ContentBlock>

Claude’s response content array. Element types match the ContentBlock descriptions above (text / image / document / tool_use / tool_result / thinking / web_search_result, etc.).

If the last message in your request has role: "assistant", this reply’s content will continue directly after that previous content, implementing prefix constraints.

Common returned block example (text-only):

{
  "content": [
    {
      "type": "text",
      "text": "Hi! My name is Claude.",
      "citations": [ ... ]
    }
  ]
}

stop_reason string

Why the model stopped generating.

Possible values:

• "end_turn": natural completion; one answer is finished;
• "max_tokens": reached max_tokens or the model limit;
• "stop_sequence": hit a sequence in stop_sequences;
• "tool_use": the response includes one or more tool_use content blocks;
• "pause_turn": for long-running server tool calls, the model pauses and waits for you to send back context to continue generating;
• "refusal": safety classifier intervened and the model refused the request.

In non-streaming mode, stop_reason is always non-null; in streaming mode, it is non-null only in certain events.

stop_sequence string | null

If stop_reason = "stop_sequence", this is the matched string; otherwise it is null.

usage object

Token usage and tool usage statistics for this request.

usage.cache_creation object

A breakdown of input tokens consumed to create new Prompt Cache entries (cache breakpoints) in this request.

• ephemeral_1h_input_tokens number
  Input tokens counted toward cache writes for newly created ephemeral cache breakpoints with a 1-hour TTL.
  Typically corresponds to parts where you set cache_control: { type: "ephemeral", ttl: "1h" }.

• ephemeral_5m_input_tokens number
  Input tokens counted toward cache writes for newly created ephemeral cache breakpoints with a 5-minute TTL.
  Typically corresponds to cache breakpoints with ttl: "5m".

Note: These fields only account for the “cache write” cost, not the “read” cost when reused later.

cache_creation_input_tokens number

The total number of input tokens written for all newly created cache breakpoints (both 5 minutes and 1 hour) in this request.

• Equals cache_creation.ephemeral_1h_input_tokens + cache_creation.ephemeral_5m_input_tokens;
• These tokens are billed as part of this request and are also written into cache for future reuse.

cache_read_input_tokens number

The number of input tokens read from existing Prompt Cache hits in this request.

• These tokens are not billed again as normal input in this request (or billed under cache pricing policy),
  and they also do not count against the context-length quota (depending on Anthropic’s internal implementation);
• Non-zero only when cache hits occur.

input_tokens number

The number of input tokens that are actually counted toward inference for this request.

output_tokens number

The number of output tokens generated by Claude in this request.

server_tool_use object

Server tool usage statistics in this request (server tools hosted by Anthropic).

• web_search_requests number The number of times the Web Search tool was actually invoked in this request.
  • Each time Claude emits a type: "server_tool_use" with name: "web_search" and it is successfully executed by the backend, this counter increments by 1;
  • Useful for tracking how many Web searches were used to obtain real-time information in the answer.

If Web Search is not enabled or not triggered in this request, this value is 0.

service_tier "standard" | string

The service tier / capacity layer actually used to process the request.

• Corresponds to your requested service_tier (e.g., "standard_only") and the system’s routing result;
• Common examples:
  • "standard": standard capacity tier;
  • Or other internal identifier strings for different service channels or priorities.

Streaming: returns multiple SSE event objects

When you set stream: true in the request, Anthropic continuously pushes a series of events via SSE (Server‑Sent Events). Each event is a JSON object. The client should:

1. Read each SSE event in arrival order;
2. Determine the event type via event: <type>;
3. Incrementally assemble the complete message from the JSON in data:.

Common event types:

• message_start
• content_block_start
• content_block_delta
• content_block_stop
• message_delta
• message_stop
• error

Top level: basic SSE event format

Each event line sent by the server looks like:

event: content_block_delta
data: { ...JSON object... }

Below is the data object structure by event type.

1. message_start event

Meaning: starts a new assistant message and provides the message’s basic metadata.

{
  "type": "message_start",
  "message": {
    "id": "msg_01ExampleID",
    "type": "message",
    "role": "assistant",
    "model": "claude-3-5-sonnet-20241022",
    "content": [],
    "stop_reason": null,
    "stop_sequence": null,
    "usage": {
      "input_tokens": 25,
      "output_tokens": 0
    }
  }
}

type string

Event type, fixed to:

• "message_start"

message object

The basic structure of the Message that is about to be streamed. Fields largely match the non-streaming Message top level (but content is usually an empty array and output_tokens may be 0 at this point):

• id string: message ID
• type string: fixed "message"
• role string: fixed "assistant"
• model string: actual model name used
• content array: initially empty; filled incrementally by subsequent content_block_* events
• stop_reason string or null: initially null; updated later via message_delta
• stop_sequence string or null: initially null; updated later via message_delta
• usage object: currently known token usage (output_tokens starts at 0; final value is provided in message_delta)

2. content_block_start event

Meaning: starts a new content block (e.g., a segment of text, or a tool invocation).

{
  "type": "content_block_start",
  "index": 0,
  "content_block": {
    "type": "text",
    "text": ""
  }
}

type string

Event type, fixed to:

• "content_block_start"

index integer

The index of this block in the overall message.content array (0-based).
Subsequent content_block_delta / content_block_stop events with the same index refer to the same block.

content_block object

The initial structure of the content block. Same as content[i] in non-streaming responses, but often an “empty shell”; the real text or parameters are added incrementally through subsequent delta events.

Typical examples:

1. Text block start:

   {
     "type": "text",
     "text": ""
   }

2. Tool invocation block start:

   {
     "type": "tool_use",
     "id": "toolu_01H...",
     "name": "get_weather",
     "input": {}
   }

3. content_block_delta event

Meaning: an “incremental update” to a content block, mainly appending text or progressively building tool invocation parameters.

{
  "type": "content_block_delta",
  "index": 0,
  "delta": {
    "type": "text_delta",
    "text": "Hello, "
  }
}

Or for tool invocation parameters:

{
  "type": "content_block_delta",
  "index": 1,
  "delta": {
    "type": "input_json_delta",
    "partial_json": "{\"location\": \"San "
  }
}

type string

Event type, fixed to:

• "content_block_delta"

index integer

The content block index this delta applies to.
Must match a previously received content_block_start.index.

delta object

The incremental payload. Its structure depends on the target block type.

Text block delta: type = "text_delta"

{
  "type": "text_delta",
  "text": "Hello, "
}

type string

• "text_delta"

text string

The new text fragment. The client should concatenate text from all text_delta events for the same index in order to form the full text.

Tool invocation parameter delta: type = "input_json_delta"

When the assistant initiates tool_use, the tool parameters input may also be assembled via multiple deltas.

{
  "type": "input_json_delta",
  "partial_json": "\"Francisco\", \"unit\": \"celsius\"}"
}

type string

• "input_json_delta"

partial_json string

A JSON fragment (string) that must be concatenated with earlier/later fragments to form the complete input object.
You should not use this parameter to call your tool until all deltas are received and the JSON is fully parsed.

4. content_block_stop event

Meaning: indicates that incremental generation for a content block has completed.

{
  "type": "content_block_stop",
  "index": 0
}

type string

Event type, fixed to:

• "content_block_stop"

index integer

The content block index. Indicates that the text or tool parameters for this index are complete and no further deltas will be sent.

5. message_delta event

Meaning: final incremental updates to the Message metadata, such as stop_reason, usage, etc.

{
  "type": "message_delta",
  "delta": {
    "stop_reason": "end_turn",
    "stop_sequence": null,
    "usage": {
      "output_tokens": 73
    }
  }
}

type string

Event type, fixed to:

• "message_delta"

delta object

Incremental updates to top-level message fields. Common fields:

stop_reason string or null

Same as stop_reason in non-streaming responses, but provided via delta only when finalized:

• "end_turn"
• "max_tokens"
• "stop_sequence"
• "tool_use"
• null

stop_sequence string or null

Used with stop_reason = "stop_sequence"; otherwise usually null.

usage object

Contains only usage fields added/updated in this delta. Most commonly:

• output_tokens integer
  The final total output token count (usually provided as a complete value in the last message_delta).

6. message_stop event

Meaning: indicates that the stream for this message has fully ended and no more events will be sent.

{
  "type": "message_stop"
}

type string

Event type, fixed to:

• "message_stop"

This event contains no other fields. After receiving it, the client can assume:

• All content_block_* events have ended;
• All message_delta updates have ended;
• The data can be assembled into a final Message object.

7. error event (in exceptional cases)

If an error occurs during the request or generation, you may receive an error event and then the stream will terminate.

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "Your request is malformed."
  }
}

type string

Event type, fixed to:

• "error"

error object

Error details.

• type string
  Error type, for example:
  • "invalid_request_error"
  • "authentication_error"
  • "rate_limit_error"
  • "api_error"
• message string
  A human-readable error description for logging and debugging.

TypeScript

import Anthropic from '@anthropic-ai/sdk';

// 1. Initialize the Anthropic client
const anthropic = new Anthropic({
  // 2. Replace with the API Key you obtained from the ZenMux user console
  apiKey: '<YOUR ZENMUX_API_KEY>', 
  // 3. Point the base URL to the ZenMux endpoint
  baseURL: "https://zenmux.ai/api/anthropic", 
});

async function main () {
    const msg = await anthropic.messages.create({
        model: "anthropic/claude-sonnet-4.5",
        max_tokens: 1024,
        messages: [{ role: "user", content: "Hello, Claude" }],
    });
    console.log(msg);
}

main();

Python

import anthropic

## 1. Initialize the Anthropic client
client = anthropic.Anthropic(
    # Replace with the API Key you obtained from the ZenMux user console
    api_key="<YOUR ZENMUX_API_KEY>", 
    # 3. Point the base URL to the ZenMux endpoint
    base_url="https://zenmux.ai/api/anthropic"
)
message = client.messages.create(
    model="anthropic/claude-sonnet-4.5",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "Hello, Claude"}
    ]
)
print(message.content)

cURL

curl https://zenmux.ai/api/anthropic/v1/messages/v1/messages \
     --header "x-api-key: $ZENMUX_API_KEY" \
     --header "anthropic-version: 2023-06-01" \
     --header "content-type: application/json" \
     --data \
'{
    "model": "anthropic/claude-sonnet-4.5",
    "max_tokens": 1024,
    "messages": [
        {"role": "user", "content": "Hello, world"}
    ]
}'

json

{
  "model": "anthropic/claude-sonnet-4.5",
  "id": "d0558ffe17be44268a7506db5f0ded62",
  "type": "message",
  "role": "assistant",
  "content": [
    {
      "type": "text",
      "text": "Hello! How can I help you today?"
    }
  ],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {
    "input_tokens": 10,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "cache_creation": {
      "ephemeral_5m_input_tokens": 0,
      "ephemeral_1h_input_tokens": 0
    },
    "output_tokens": 12,
    "service_tier": "standard"
  }
}