Claude Haiku 4.5

anthropic/claude-haiku-4-5 — a llm model by Anthropic. Call it through the gateway’s standard OpenAI-compatible door using the catalog slug; the full parameter schema is reproduced in full below.

Catalog slug	anthropic/claude-haiku-4-5
Kind	llm
Vendor	Anthropic
Context window	200,000 tokens
Native model id	claude-haiku-4-5

How to call

Through the Sociaro SDK / OpenAI-compatible door

This is an Anthropic model. When you call chat.completions.create, the SDK automatically translates the request to the Anthropic Messages format and sends it to POST /v1/messages. Calling the gateway directly (without the SDK) must use POST /v1/messages — POST /v1/chat/completions returns 400 for Anthropic models (that door serves OpenAI-format models only).

from sociaro_ai import Sociaro

client = Sociaro(api_key="gw_live_...")   # base_url defaults to https://api.sociaro.com

resp = client.chat.completions.create(
    model="anthropic/claude-haiku-4-5",          # always the catalog slug
    messages=[{"role": "user", "content": "Hello"}],
    # any parameter from the schema below is passed through unchanged
)
print(resp.choices[0].message.content)

TypeScript: client.chat.completions.create({ model: "anthropic/claude-haiku-4-5", messages: [...] }) — see the TypeScript SDK.

Through Anthropic’s native API (advanced)

Optional: send Anthropic’s native request to the /anthropic prefix, using the native model id claude-haiku-4-5. The gateway injects the upstream Anthropic credential and forwards the body unchanged (catalog, pricing and entitlements still apply).

curl https://api.sociaro.com/anthropic/v1/messages \
  -H "x-api-key: $SOCIARO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "model": "claude-haiku-4-5", ... }'

The exact path, request body and response shape are documented in full in the schema below.

Parameters — full schema

The complete, one-to-one API schema for this model, transcribed from the official source. Nothing omitted. Examples in the schema may use a representative model of the family — for this model use the slug anthropic/claude-haiku-4-5 (or the native id claude-haiku-4-5 for the native API) as shown in How to call above.

Coverage map (model enum)

The model field is an enum, and free-string dated aliases are also accepted.

Model id	Tagline
claude-fable-5	Next generation of intelligence for the hardest knowledge work and coding problems
claude-mythos-5	Most capable model for cybersecurity and biology research
claude-mythos-preview	New class of intelligence, strongest in coding and cybersecurity
claude-opus-4-8	Frontier intelligence for long-running agents and coding
claude-opus-4-7	Frontier intelligence for long-running agents and coding
claude-opus-4-6	Frontier intelligence for long-running agents and coding
claude-opus-4-5, claude-opus-4-5-20251101	Premium model combining maximum intelligence with practical performance
claude-opus-4-1, claude-opus-4-1-20250805	Exceptional model for specialized complex tasks
claude-opus-4-0, claude-opus-4-20250514	Powerful model for complex tasks
claude-sonnet-4-6	Best combination of speed and intelligence
claude-sonnet-4-5, claude-sonnet-4-5-20250929	High-performance model for agents and coding
claude-sonnet-4-0, claude-sonnet-4-20250514	High-performance model with extended thinking
claude-haiku-4-5, claude-haiku-4-5-20251001	Fastest model with near-frontier intelligence

Additional dated/legacy aliases (e.g. claude-3-haiku-20240307) are accepted as free strings. For gateway slugs anthropic/claude-* the native prefix is transparent: these parameters pass through to /v1/messages one-to-one.

Critical per-model differences

1. Sampling params are deprecated. temperature, top_p, top_k are marked Deprecated. Models released after Claude Opus 4.6 do not support setting them:

   • temperature: only 1.0 is accepted (backwards compatibility); any other value returns HTTP 400.
   • top_p: only >= 0.99 is accepted (backwards compatibility); any other value returns HTTP 400.
   • top_k: any value is rejected with HTTP 400.

2. Extended thinking mode varies by model (source: extended-thinking guide):

   • claude-fable-5, claude-mythos-5: extended thinking is always on and cannot be disabled. Manual thinking:{type:"enabled", budget_tokens:N} is not supported (HTTP 400), and thinking:{type:"disabled"} returns an error. Use adaptive thinking and control depth with output_config.effort.
   • claude-opus-4-8, claude-opus-4-7: manual thinking:{type:"enabled"} is not supported (HTTP 400). Use thinking:{type:"adaptive"} with output_config.effort; the model decides whether and how much to think.
   • claude-mythos-preview: adaptive is the default; type:"enabled" is also accepted; type:"disabled" is not supported. display defaults to "omitted" here (not "summarized") — pass display:"summarized" to receive summaries.
   • claude-opus-4-6, claude-sonnet-4-6: adaptive thinking is recommended; manual type:"enabled" (including interleaved) is deprecated but still functional.
   • Other current models (haiku-4-5, sonnet-4-5/4-0, opus-4-5/4-1/4-0): manual extended thinking (type:"enabled", budget_tokens:N) is supported. In summary, manual thinking is supported on all current models except Fable 5, Mythos 5, Opus 4.8, and Opus 4.7, where it returns HTTP 400.

3. output_config.effort controls reasoning depth for adaptive-thinking models (Opus 4.8/4.7, Fable 5, Mythos 5, Mythos Preview); pairs with thinking:{type:"adaptive"}.

4. Priority Tier (service_tier) is supported on all models except claude-mythos-preview and claude-mythos-5.

5. max_tokens maximum differs per model (see the Models overview). 0 is legal (warms the prompt cache without generating).

---

Endpoint

• Create: POST https://api.anthropic.com/v1/messages
• Token-counting twin: POST https://api.anthropic.com/v1/messages/count_tokens — same body params (no sampling/output-generation effect); returns { "input_tokens": N }.
• Send a structured list of input messages (text and/or image/document content); the model generates the next message. Usable for single queries or stateless multi-turn conversations.

Request HTTP headers

Header	Required	Value / notes
x-api-key	yes	Your Anthropic API key.
anthropic-version	yes	API version, e.g. 2023-06-01. Send it on every request.
content-type	yes	application/json.
anthropic-beta	no	Comma-separated beta feature flags (fine-grained tool streaming, extended cache TTL, etc.).

When service_tier="auto" and a model has a Priority Tier commitment, the response also returns rate headers: anthropic-priority-input-tokens-limit/-remaining/-reset and anthropic-priority-output-tokens-limit/-remaining/-reset.

---

Request body — top-level parameters

Name	Type	Required	Default	Allowed / range	Description
model	Model (string enum or free-string alias)	yes	—	See coverage map enum	The model that will complete your prompt.
messages	array of MessageParam {content, role}	yes	—	—	Input messages. Models operate on alternating user/assistant turns; consecutive same-role turns are merged. A trailing assistant turn is continued (prefill).
max_tokens	integer	yes	—	minimum 0 (model-specific maximum)	Max tokens to generate before stopping. The model may stop earlier. 0 warms the prompt cache without generating a response. Different models have different maximums.
system	string OR array of TextBlockParam {text, type, cache_control, citations}	no	—	—	System prompt (context/instructions/role). There is no "system" role in messages; use this top-level field.
stop_sequences	array of string	no	—	—	Custom text sequences that stop generation. On match, response stop_reason = "stop_sequence" and stop_sequence = the matched string.
stream	boolean	no	false	true/false	Incrementally stream the response via server-sent events.
temperature	number	no	1.0	0.0–1.0	Deprecated. Amount of randomness. Models released after Opus 4.6 do not support setting it — only 1.0 is accepted for backwards compatibility; other values return HTTP 400. Even 0.0 is not fully deterministic.
top_p	number	no	—	0.0–1.0	Deprecated. Nucleus sampling. Models released after Opus 4.6 reject it — only >= 0.99 is accepted for backwards compatibility; other values return HTTP 400. Advanced use only.
top_k	number	no	—	minimum 0	Deprecated. Sample only from top-K tokens. Models released after Opus 4.6 reject any value with HTTP 400. Advanced use only.
tools	array of ToolUnion	no	—	—	Tool definitions the model may call (tool_use blocks). Includes custom client tools and server tools. See shapes below.
tool_choice	ToolChoice (union)	no	auto (when tools present)	{type:"auto"|"any"|"tool"|"none"}	How the model uses tools. auto = model decides; any = must use some tool; tool (+name) = must use the named tool; none = no tools. auto/any/tool accept optional disable_parallel_tool_use (boolean, default false).
thinking	ThinkingConfigParam (union)	no	—	enabled / disabled / adaptive (see below)	Configuration for extended thinking. See the per-model matrix above and the dedicated section below.
metadata	Metadata {user_id}	no	—	user_id maxLength 512	Metadata about the request. user_id: opaque external user id (uuid/hash) for abuse detection — must not contain name/email/phone.
service_tier	string (enum)	no	"auto"	"auto", "standard_only"	Whether to use priority capacity (if available) or standard. auto may use priority then fall back; standard_only forces standard.
output_config	OutputConfig {effort?, format?}	no	—	see below	Output options: reasoning effort and/or structured-output JSON schema.
container	string	no	—	—	Container identifier for reuse across requests (code execution tool).
inference_geo	string	no	workspace default_inference_geo	—	Geographic region for inference processing.

MCP is not configured through a top-level request field on this endpoint; there is no mcp_servers parameter in the request body.

thinking — ThinkingConfigParam (discriminated union)

type discriminates three variants. display is present only on enabled and adaptive; it is absent on disabled (ThinkingConfigDisabled properties = {type} only).

Variant	Shape	Fields
ThinkingConfigEnabled	{type:"enabled", budget_tokens, display?}	type="enabled". budget_tokens (number, minimum 1024, must be < max_tokens; counts toward max_tokens). display: "summarized" (default) or "omitted".
ThinkingConfigDisabled	{type:"disabled"}	type="disabled" only. No display field.
ThinkingConfigAdaptive	{type:"adaptive", display?}	type="adaptive". display: "summarized" (default) or "omitted". Model picks its own budget.

display semantics: "summarized" → thinking returned normally; "omitted" → thinking content redacted but a signature is returned for multi-turn continuity. Defaults to "summarized" (except claude-mythos-preview, where it defaults to "omitted").

output_config — OutputConfig

Field	Type	Required	Allowed	Notes
effort	string enum	no	"low", "medium", "high", "xhigh", "max"	Reasoning effort level (adaptive-thinking models).
format	JSONOutputFormat {schema, type}	no	type="json_schema"; schema=JSON Schema	Structured-outputs JSON schema for Claude’s output format.

tool_choice — ToolChoice (union)

Variant	Shape
ToolChoiceAuto	{type:"auto", disable_parallel_tool_use?} — model decides; disable_parallel_tool_use default false, if true outputs at most one tool use.
ToolChoiceAny	{type:"any", disable_parallel_tool_use?} — must use some tool; if disable_parallel_tool_use=true, outputs exactly one tool use.
ToolChoiceTool	{name, type:"tool", disable_parallel_tool_use?} — name (string) = the tool to use; if disable_parallel_tool_use=true, exactly one tool use.
ToolChoiceNone	{type:"none"} — model not allowed to use tools.

cache_control — CacheControlEphemeral

Applied per-block (text/image/document/tool/tool_result/system blocks) to set a prompt-cache breakpoint.

Field	Type	Required	Allowed	Default
type	string	yes	"ephemeral"	—
ttl	string	no	"5m" (5 minutes), "1h" (1 hour)	"5m"

---

messages[].content — block types (request side, ContentBlockParam)

content is a string (shorthand for one text block) OR an array of typed blocks. Each input message: { "role": "user"|"assistant", "content": ... }. Block type values:

Block type	Key fields	Notes
text	text (string); optional cache_control, citations (array of TextCitationParam)	Plain text.
image	source: base64 {data, media_type} (image/jpeg/png/gif/webp), or url, or file	Vision input.
document	source: base64 PDF {data, media_type:"application/pdf"}, url, file, text, or content; optional citations {enabled}, title, context, cache_control	PDF / document input.
tool_use	id, name, input	Model-generated tool call (assistant turns).
tool_result	tool_use_id, content (string or blocks), is_error?, cache_control?	Result returned to the model.
thinking / redacted_thinking	thinking/data, signature	Extended-thinking blocks — echo back for multi-turn continuity.
search_result	content, source, title; optional citations {enabled}, cache_control	Search-result content.
server_tool_use, web_search_tool_result, web_fetch_tool_result, code_execution_tool_result, bash_code_execution_tool_result, text_editor_code_execution_tool_result, container_upload	varies	Server-tool / container blocks.

Most blocks accept optional cache_control: {type:"ephemeral", ttl:"5m"|"1h"}.

---

tools[] item shape — custom client tool (Tool)

Field	Type	Required	Constraints / values	Notes
name	string	yes	minLength 1, maxLength 128	How the model calls the tool (also appears in tool_use blocks).
input_schema	object {type:"object", properties?, required?}	yes	JSON Schema (draft 2020-12); properties = object map; required = array of string	Shape of the tool input.
description	string	no	—	Strongly recommended; more detail → better tool-use quality.
type	string	no	"custom"	Custom tool marker.
cache_control	CacheControlEphemeral {type, ttl?}	no	ttl "5m"/"1h"	Cache breakpoint at this tool.
allowed_callers	array of string	no	"direct", "code_execution_20250825", "code_execution_20260120"	Who may call the tool.
defer_loading	boolean	no	—	If true, tool loaded lazily via tool-search tool_reference.
eager_input_streaming	boolean	no	—	Stream tool input incrementally (fine-grained streaming).
input_examples	array	no	—	Example inputs.
strict	boolean	no	—	Guarantee schema validation on tool names/inputs.

Server-tool variants (ToolUnion, discriminated by type with a fixed dated version)

Each has its own fixed name and config; some support strict, cache_control, allowed_callers, and a user_location {city?, country?, region?, timezone?} (web search/fetch):

• Bash: bash_20250124
• Code execution: code_execution_20250522, code_execution_20250825, code_execution_20260120
• Web search: web_search_20250305, web_search_20260209
• Web fetch: web_fetch_20250910, web_fetch_20260209, web_fetch_20260309
• Text editor: text_editor_20250124, text_editor_20250429, text_editor_20250728 (plus the text_editor_code_execution result block family)
• Memory: memory_20250818
• Tool search / clear-tool-uses: dated context-management tool variants (tool_search, with tool_reference for deferred-loaded tools)

---

Response body (HTTP 200) — Message object

Message {id, container, content, model, role, stop_reason, stop_sequence, stop_details, type, usage}.

Field	Type	Notes
id	string	Unique object id (e.g. msg_…). Format/length may change.
container	Container {id, expires_at} or null	Code-execution container info: id (string), expires_at (string, expiry time).
content	array of ContentBlock	Generated content (text, tool_use, thinking, server-tool blocks, …). Continues from a trailing assistant prefill if present.
model	string (Model enum)	The model that handled the request.
role	"assistant"	Always "assistant".
stop_reason	StopReason enum or null	"end_turn", "max_tokens", "stop_sequence", "tool_use", "pause_turn", "refusal". Non-null in non-streaming; null only in message_start for streaming.
stop_sequence	string or null	Matched custom stop sequence, if any.
stop_details	RefusalStopDetails {category, explanation, type} or null	On refusal. category: "cyber"/"bio"/"reasoning_extraction" or null (when the refusal doesn’t map to a named category). explanation: string (not guaranteed stable; null if unavailable). type: "refusal".
type	"message"	Always "message".
usage	Usage object	Billing / rate-limit token counts (below).

stop_reason values (StopReason)

• "end_turn" — model reached a natural stopping point.
• "max_tokens" — exceeded requested max_tokens or the model’s maximum.
• "stop_sequence" — one of your custom stop_sequences was generated.
• "tool_use" — the model invoked one or more tools.
• "pause_turn" — a long-running turn was paused; pass the response back as-is in a subsequent request to continue.
• "refusal" — streaming classifiers intervened for a potential policy violation.

Response content blocks (ContentBlock)

Each has a type. Includes:

• TextBlock {citations, text, type:"text"}. citations is an array; citation variants depend on cited document type:
  • CitationCharLocation {cited_text, document_index, document_title, end_char_index, file_id, start_char_index, type:"char_location"} (plain text)
  • CitationPageLocation {cited_text, document_index, document_title, end_page_number, file_id, start_page_number, type:"page_location"} (PDF)
  • CitationContentBlockLocation {cited_text, document_index, document_title, end_block_index, file_id, start_block_index, type:"content_block_location"} (content document)
  • CitationsWebSearchResultLocation {cited_text, encrypted_index, title, type:"web_search_result_location", url}
  • CitationsSearchResultLocation {cited_text, end_block_index, search_result_index, source, start_block_index, title, type:"search_result_location"}
• ToolUseBlock {id, name, input, type:"tool_use"}
• ThinkingBlock {thinking, signature, type:"thinking"} / RedactedThinkingBlock {data, type:"redacted_thinking"}
• Server-tool / container result blocks (server_tool_use, web_search_tool_result, web_fetch_tool_result, code_execution_tool_result, container_upload, …).

usage object (Usage)

Total input tokens = input_tokens + cache_creation_input_tokens + cache_read_input_tokens. output_tokens is non-zero even for empty output and is the authoritative billing total.

Field	Type	Notes
input_tokens	number	Non-cached input tokens used.
output_tokens	number	Output tokens used. Inclusive, authoritative billing total.
cache_creation_input_tokens	number	Tokens used to create the cache entry.
cache_read_input_tokens	number	Tokens read from the cache.
cache_creation	CacheCreation {ephemeral_1h_input_tokens, ephemeral_5m_input_tokens}	Cached-token breakdown by TTL.
output_tokens_details	OutputTokensDetails {thinking_tokens}	thinking_tokens = output tokens spent on internal reasoning (including thinking-block delimiter tokens); read-only decomposition, ≤ output_tokens. Reflects raw reasoning, not the possibly-shorter summarized text.
server_tool_use	ServerToolUsage {web_fetch_requests, web_search_requests}	Server-tool request counts.
service_tier	enum "standard" / "priority" / "batch"	The tier actually used.
inference_geo	string	Region where inference ran.

---

Streaming (when stream: true)

Server-sent events; RawMessageStreamEvent is one of: message_start (carries the initial Message with null stop_reason), message_delta ({delta, type, usage} — cumulative usage, top-level stop_reason/stop_sequence deltas), message_stop, content_block_start, content_block_delta, content_block_stop, plus ping and error events.

---

Minimal request example

curl https://api.anthropic.com/v1/messages \
  -H "content-type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -d '{
    "model": "claude-sonnet-4-6",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Hello, Claude"}
    ]
  }'

Full request example — adaptive thinking + structured output + tools

curl https://api.anthropic.com/v1/messages \
  -H 'content-type: application/json' \
  -H 'anthropic-version: 2023-06-01' \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  --max-time 600 \
  -d '{
    "model": "claude-opus-4-8",
    "max_tokens": 4096,
    "system": [{"type": "text", "text": "Today is 2026-06-01."}],
    "messages": [
      {"role": "user", "content": "What is the weather in SF? Reply as JSON."}
    ],
    "thinking": { "type": "adaptive", "display": "summarized" },
    "output_config": {
      "effort": "high",
      "format": {
        "type": "json_schema",
        "schema": {
          "type": "object",
          "properties": {"city": {"type": "string"}, "temp_c": {"type": "number"}},
          "required": ["city", "temp_c"]
        }
      }
    },
    "tools": [
      {
        "name": "get_weather",
        "description": "Get current weather for a location.",
        "input_schema": {
          "type": "object",
          "properties": {"location": {"type": "string"}, "unit": {"type": "string"}},
          "required": ["location"]
        }
      }
    ],
    "tool_choice": { "type": "auto", "disable_parallel_tool_use": false },
    "service_tier": "auto",
    "metadata": { "user_id": "9f2b...opaque-hash" }
  }'

For opus-4-8, opus-4-7, fable-5, and mythos-5, manual thinking:{type:"enabled", budget_tokens:N} returns HTTP 400 — use type:"adaptive" with output_config.effort as above. For sonnet-4-6 and opus-4-6, manual enabled still works but is deprecated. For haiku-4-5 and the 4-5/4-1/4-0 generations, manual enabled is fully supported. temperature/top_p/top_k are deprecated and rejected on models newer than Opus 4.6 (except the backwards-compatibility values temperature=1.0 and top_p>=0.99).

Response example (HTTP 200)

{
  "id": "msg_013Zva2CMHLNnXjNJJKqJ2EF",
  "container": { "id": "id", "expires_at": "2019-12-27T18:11:19.117Z" },
  "content": [
    {
      "citations": [
        {
          "cited_text": "cited_text",
          "document_index": 0,
          "document_title": "document_title",
          "end_char_index": 0,
          "file_id": "file_id",
          "start_char_index": 0,
          "type": "char_location"
        }
      ],
      "text": "Hi! My name is Claude.",
      "type": "text"
    }
  ],
  "model": "claude-opus-4-6",
  "role": "assistant",
  "stop_details": { "category": "cyber", "explanation": "explanation", "type": "refusal" },
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "type": "message",
  "usage": {
    "cache_creation": { "ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0 },
    "cache_creation_input_tokens": 2051,
    "cache_read_input_tokens": 2051,
    "inference_geo": "inference_geo",
    "input_tokens": 2095,
    "output_tokens": 503,
    "output_tokens_details": { "thinking_tokens": 0 },
    "server_tool_use": { "web_fetch_requests": 2, "web_search_requests": 0 },
    "service_tier": "standard"
  }
}

Token-counting twin response example

{ "input_tokens": 2095 }