Blog
跳至主要內容

第 0 天支援：Claude 4.5 Opus（+進階功能）

Sameer Kankute
Sameer Kankute
SWE @ LiteLLM (LLM Translation)
Krrish Dholakia
Krrish Dholakia
CEO, LiteLLM
Ishaan Jaffer
Ishaan Jaffer
CTO, LiteLLM

本指南涵蓋 Anthropic 的最新模型（Claude Opus 4.5）及其目前可在 LiteLLM 中使用的進階功能：Tool Search、Programmatic Tool Calling、Tool Input Examples 與 Effort 參數。

功能支援的模型
Tool SearchClaude Opus 4.5, Sonnet 4.5
Programmatic Tool CallingClaude Opus 4.5, Sonnet 4.5
Input ExamplesClaude Opus 4.5, Sonnet 4.5
Effort 參數僅 Claude Opus 4.5

支援的提供者：Anthropic、Bedrock、Vertex AI、Azure AI。

使用​

import os
from litellm import completion

# set env - [OPTIONAL] replace with your anthropic key
os.environ["ANTHROPIC_API_KEY"] = "your-api-key"

messages = [{"role": "user", "content": "Hey! how's it going?"}]

## OPENAI /chat/completions API format
response = completion(model="claude-opus-4-5-20251101", messages=messages)
print(response)

使用 - Bedrock​

資訊

LiteLLM 使用 boto3 函式庫與 Bedrock 進行驗證。

如需更多與 Bedrock 驗證的方式，請參閱 Bedrock 文件。

import os
from litellm import completion

os.environ["AWS_ACCESS_KEY_ID"] = ""
os.environ["AWS_SECRET_ACCESS_KEY"] = ""
os.environ["AWS_REGION_NAME"] = ""

## OPENAI /chat/completions API format
response = completion(
  model="bedrock/us.anthropic.claude-opus-4-5-20251101-v1:0",
  messages=[{ "content": "Hello, how are you?","role": "user"}]
)

使用 - Vertex AI​

from litellm import completion
import json 

## GET CREDENTIALS 
## RUN ## 
# !gcloud auth application-default login - run this to add vertex credentials to your env
## OR ## 
file_path = 'path/to/vertex_ai_service_account.json'

# Load the JSON file
with open(file_path, 'r') as file:
    vertex_credentials = json.load(file)

# Convert to JSON string
vertex_credentials_json = json.dumps(vertex_credentials)

## COMPLETION CALL 
response = completion(
  model="vertex_ai/claude-opus-4-5@20251101",
  messages=[{ "content": "Hello, how are you?","role": "user"}],
  vertex_credentials=vertex_credentials_json,
  vertex_project="your-project-id",
  vertex_location="us-east5"
)

使用 - Azure Anthropic（Azure Foundry Claude）​

LiteLLM 會透過 azure_ai/ 提供者將 Azure Claude 部署導向，讓 Azure Foundry 上的 Claude Opus 模型可繼續使用 Tool Search、Effort、串流以及其餘進階功能組合。將 AZURE_AI_API_BASE 指向 https://<resource>.services.ai.azure.com/anthropic（LiteLLM 會自動附加 /v1/messages），並使用 AZURE_AI_API_KEY 或 Azure AD 權杖進行驗證。

import os
from litellm import completion

# Configure Azure credentials
os.environ["AZURE_AI_API_KEY"] = "your-azure-ai-api-key"
os.environ["AZURE_AI_API_BASE"] = "https://my-resource.services.ai.azure.com/anthropic"

response = completion(
    model="azure_ai/claude-opus-4-1",
    messages=[{"role": "user", "content": "Explain how Azure Anthropic hosts Claude Opus differently from the public Anthropic API."}],
    max_tokens=1200,
    temperature=0.7,
    stream=True,
)

for chunk in response:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

這可讓 Claude 透過按需動態載入工具，而不是一開始就將所有工具載入到 context window 中，進而支援數千個工具。

使用範例​

import litellm
import os

# Configure your API key
os.environ["ANTHROPIC_API_KEY"] = "your-api-key"

# Define your tools with defer_loading
tools = [
    # Tool search tool (regex variant)
    {
        "type": "tool_search_tool_regex_20251119",
        "name": "tool_search_tool_regex"
    },
    # Deferred tools - loaded on-demand
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get the current weather in a given location. Returns temperature and conditions.",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": "The city and state, e.g. San Francisco, CA"
                    },
                    "unit": {
                        "type": "string",
                        "enum": ["celsius", "fahrenheit"],
                        "description": "Temperature unit"
                    }
                },
                "required": ["location"]
            }
        },
        "defer_loading": True  # Load on-demand
    },
    {
        "type": "function",
        "function": {
            "name": "search_files",
            "description": "Search through files in the workspace using keywords",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {"type": "string"},
                    "file_types": {
                        "type": "array",
                        "items": {"type": "string"}
                    }
                },
                "required": ["query"]
            }
        },
        "defer_loading": True
    },
    {
        "type": "function",
        "function": {
            "name": "query_database",
            "description": "Execute SQL queries against the database",
            "parameters": {
                "type": "object",
                "properties": {
                    "sql": {"type": "string"}
                },
                "required": ["sql"]
            }
        },
        "defer_loading": True
    }
]

# Make a request - Claude will search for and use relevant tools
response = litellm.completion(
    model="anthropic/claude-opus-4-5-20251101",
    messages=[{
        "role": "user",
        "content": "What's the weather like in San Francisco?"
    }],
    tools=tools
)

print("Claude's response:", response.choices[0].message.content)
print("Tool calls:", response.choices[0].message.tool_calls)

# Check tool search usage
if hasattr(response.usage, 'server_tool_use'):
    print(f"Tool searches performed: {response.usage.server_tool_use.tool_search_requests}")

針對自然語言查詢，而非 regex 模式：

tools = [
    {
        "type": "tool_search_tool_bm25_20251119",  # Natural language variant
        "name": "tool_search_tool_bm25"
    },
    # ... your deferred tools
]

程式化工具呼叫​

Programmatic tool calling 讓 Claude 能撰寫程式碼，以程式化方式呼叫您的工具。深入了解

import litellm
import json

# Define tools that can be called programmatically
tools = [
    # Code execution tool (required for programmatic calling)
    {
        "type": "code_execution_20250825",
        "name": "code_execution"
    },
    # Tool that can be called from code
    {
        "type": "function",
        "function": {
            "name": "query_database",
            "description": "Execute a SQL query against the sales database. Returns a list of rows as JSON objects.",
            "parameters": {
                "type": "object",
                "properties": {
                    "sql": {
                        "type": "string",
                        "description": "SQL query to execute"
                    }
                },
                "required": ["sql"]
            }
        },
        "allowed_callers": ["code_execution_20250825"]  # Enable programmatic calling
    }
]

# First request
response = litellm.completion(
    model="anthropic/claude-sonnet-4-5-20250929",
    messages=[{
        "role": "user",
        "content": "Query sales data for West, East, and Central regions, then tell me which had the highest revenue"
    }],
    tools=tools
)

print("Claude's response:", response.choices[0].message)

# Handle tool calls
messages = [
    {"role": "user", "content": "Query sales data for West, East, and Central regions, then tell me which had the highest revenue"},
    {"role": "assistant", "content": response.choices[0].message.content, "tool_calls": response.choices[0].message.tool_calls}
]

# Process each tool call
for tool_call in response.choices[0].message.tool_calls:
    # Check if it's a programmatic call
    if hasattr(tool_call, 'caller') and tool_call.caller:
        print(f"Programmatic call to {tool_call.function.name}")
        print(f"Called from: {tool_call.caller}")
    
    # Simulate tool execution
    if tool_call.function.name == "query_database":
        args = json.loads(tool_call.function.arguments)
        # Simulate database query
        result = json.dumps([
            {"region": "West", "revenue": 150000},
            {"region": "East", "revenue": 180000},
            {"region": "Central", "revenue": 120000}
        ])
        
        messages.append({
            "role": "user",
            "content": [{
                "type": "tool_result",
                "tool_use_id": tool_call.id,
                "content": result
            }]
        })

# Get final response
final_response = litellm.completion(
    model="anthropic/claude-sonnet-4-5-20250929",
    messages=messages,
    tools=tools
)

print("\nFinal answer:", final_response.choices[0].message.content)

工具輸入範例​

您現在可以提供 Claude 如何使用工具的範例。深入了解

import litellm

tools = [
    {
        "type": "function",
        "function": {
            "name": "create_calendar_event",
            "description": "Create a new calendar event with attendees and reminders",
            "parameters": {
                "type": "object",
                "properties": {
                    "title": {"type": "string"},
                    "start_time": {
                        "type": "string",
                        "description": "ISO 8601 format: YYYY-MM-DDTHH:MM:SS"
                    },
                    "duration_minutes": {"type": "integer"},
                    "attendees": {
                        "type": "array",
                        "items": {
                            "type": "object",
                            "properties": {
                                "email": {"type": "string"},
                                "optional": {"type": "boolean"}
                            }
                        }
                    },
                    "reminders": {
                        "type": "array",
                        "items": {
                            "type": "object",
                            "properties": {
                                "minutes_before": {"type": "integer"},
                                "method": {"type": "string", "enum": ["email", "popup"]}
                            }
                        }
                    }
                },
                "required": ["title", "start_time", "duration_minutes"]
            }
        },
        # Provide concrete examples
        "input_examples": [
            {
                "title": "Team Standup",
                "start_time": "2025-01-15T09:00:00",
                "duration_minutes": 30,
                "attendees": [
                    {"email": "alice@company.com", "optional": False},
                    {"email": "bob@company.com", "optional": False}
                ],
                "reminders": [
                    {"minutes_before": 15, "method": "popup"}
                ]
            },
            {
                "title": "Lunch Break",
                "start_time": "2025-01-15T12:00:00",
                "duration_minutes": 60
                # Demonstrates optional fields can be omitted
            }
        ]
    }
]

response = litellm.completion(
    model="anthropic/claude-sonnet-4-5-20250929",
    messages=[{
        "role": "user",
        "content": "Schedule a team meeting for tomorrow at 2pm for 45 minutes with john@company.com and sarah@company.com"
    }],
    tools=tools
)

print("Tool call:", response.choices[0].message.tool_calls[0].function.arguments)

Effort 參數：控制 Token 使用量​

使用 reasoning_effort 參數控制 Claude 在回應中投入多少 effort。這可讓您在回應完整度與 token 效率之間取得取捨。

資訊

LiteLLM 會自動將 reasoning_effort 對應到 Anthropic 的 output_config 格式，並為 Claude Opus 4.5 加上所需的 effort-2025-11-24 beta 標頭。

reasoning_effort 參數的可能值："high"、"medium"、"low"。

使用範例​

import litellm

message = "Analyze the trade-offs between microservices and monolithic architectures"

# High effort (default) - Maximum capability
response_high = litellm.completion(
    model="anthropic/claude-opus-4-5-20251101",
    messages=[{"role": "user", "content": message}],
    reasoning_effort="high"
)

print("High effort response:")
print(response_high.choices[0].message.content)
print(f"Tokens used: {response_high.usage.completion_tokens}\n")

# Medium effort - Balanced approach
response_medium = litellm.completion(
    model="anthropic/claude-opus-4-5-20251101",
    messages=[{"role": "user", "content": message}],
    reasoning_effort="medium"
)

print("Medium effort response:")
print(response_medium.choices[0].message.content)
print(f"Tokens used: {response_medium.usage.completion_tokens}\n")

# Low effort - Maximum efficiency
response_low = litellm.completion(
    model="anthropic/claude-opus-4-5-20251101",
    messages=[{"role": "user", "content": message}],
    reasoning_effort="low"
)

print("Low effort response:")
print(response_low.choices[0].message.content)
print(f"Tokens used: {response_low.usage.completion_tokens}\n")

# Compare token usage
print("Token Comparison:")
print(f"High:   {response_high.usage.completion_tokens} tokens")
print(f"Medium: {response_medium.usage.completion_tokens} tokens")
print(f"Low:    {response_low.usage.completion_tokens} tokens")
🚅
LiteLLM Enterprise
為正式環境打造的 SSO/SAML、稽核記錄、支出追蹤、多團隊管理與防護欄。
深入瞭解 →