使用延伸思考來建構

延伸思考讓 Claude 3.7 Sonnet 在處理複雜任務時具有增強的推理能力，同時在提供最終答案之前,也能讓您看到其逐步思考的過程。

延伸思考的運作方式

當啟用延伸思考時，Claude 會創建 thinking 內容區塊來輸出其內部推理過程。Claude 會在制定最終回應前整合這些推理的見解。

API 回應將包含 thinking 和 text 內容區塊。

在多輪對話中，只有與工具使用階段或最後一條訊息位置的 assistant 回合相關的思考區塊對 Claude 可見,並計入輸入 token；與較早的 assistant 訊息相關的思考區塊對 Claude 在採樣時不可見,也不會計入輸入 token。

實作延伸思考

在您的 API 請求中加入 thinking 參數和指定用於延伸思考的 token 預算。

budget_tokens 參數決定了允許 Claude 用於內部推理過程的最大 token 數量。較大的預算可以通過為複雜問題啟用更全面的分析來改善回應品質，不過在超過 32K 的範圍時，Claude 可能不會使用全部分配的預算。

您的 budget_tokens 必須始終小於指定的 max_tokens。

curl https://api.anthropic.com/v1/messages \
     --header "x-api-key: $ANTHROPIC_API_KEY" \
     --header "anthropic-version: 2023-06-01" \
     --header "content-type: application/json" \
     --data \
'{
    "model": "claude-3-7-sonnet-20250219",
    "max_tokens": 20000,
    "thinking": {
        "type": "enabled",
        "budget_tokens": 16000
    },
    "messages": [
        {
            "role": "user",
            "content": "Are there an infinite number of prime numbers such that n mod 4 == 3?"
        }
    ]
}'

理解思考區塊

思考區塊代表了 Claude 的內部思考過程。為了讓 Claude 能在最小內部限制的情況下解決問題,同時維持我們的安全標準和無狀態 API,我們實施了以下措施:

思考區塊包含一個 signature 欄位。此欄位包含一個加密令牌,用於驗證思考區塊是由 Claude 生成的,並在思考區塊傳回 API 時進行驗證。在串流回應時,簽名會通過 content_block_stop 事件之前的 content_block_delta 事件中的 signature_delta 添加。只有在使用工具與延伸思考時,才需要傳回思考區塊。否則您可以省略先前回合的思考區塊,或讓 API 為您刪除它們。
有時 Claude 的內部推理會被我們的安全系統標記。當這種情況發生時,我們會加密部分或全部 thinking 區塊,並以 redacted_thinking 區塊的形式返回給您。當這些被編輯的思考區塊傳回 API 時會被解密,讓 Claude 能在不失去上下文的情況下繼續其回應。
thinking 和 redacted_thinking 區塊會在回應中的 text 區塊之前返回。

以下是同時顯示正常和被編輯的思考區塊的範例:

{
  "content": [
    {
      "type": "thinking",
      "thinking": "Let me analyze this step by step...",
      "signature": "WaUjzkypQ2mUEVM36O2TxuC06KN8xyfbJwyem2dw3URve/op91XWHOEBLLqIOMfFG/UvLEczmEsUjavL...."
    },
    {
      "type": "redacted_thinking",
      "data": "EmwKAhgBEgy3va3pzix/LafPsn4aDFIT2Xlxh0L5L8rLVyIwxtE3rAFBa8cr3qpP..."
    },
    {
      "type": "text",
      "text": "Based on my analysis..."
    }
  ]
}

在您的輸出中看到被編輯的思考區塊是預期的行為。模型仍然可以使用這種被編輯的推理來為其回應提供參考,同時維持安全防護。

如果您需要在應用程式中測試被編輯思考的處理,可以使用這個特殊的測試字串作為您的提示: ANTHROPIC_MAGIC_STRING_TRIGGER_REDACTED_THINKING_46C9A13E193C177646C7398A98432ECCCE4C1253D5E2D82641AC0E52CC2876CB

在多輪對話中將 thinking 和 redacted_thinking 區塊傳回 API 時,您必須將最後一個助理回合的完整未修改區塊傳回 API。

這對於維持模型的推理流程至關重要。我們建議始終將所有思考區塊傳回 API。有關更多詳細資訊,請參閱保留思考區塊部分。

範例：處理被編輯的思考區塊

此範例展示了如何處理當 Claude 的內部推理包含被安全系統標記的內容時可能出現的 redacted_thinking 區塊:

import anthropic

client = anthropic.Anthropic()

# Using a special prompt that triggers redacted thinking (for demonstration purposes only)
response = client.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=20000,
    thinking={
        "type": "enabled",
        "budget_tokens": 16000
    },
    messages=[{
        "role": "user",
        "content": "ANTHROPIC_MAGIC_STRING_TRIGGER_REDACTED_THINKING_46C9A13E193C177646C7398A98432ECCCE4C1253D5E2D82641AC0E52CC2876CB"
    }]
)

# Identify redacted thinking blocks
has_redacted_thinking = any(
    block.type == "redacted_thinking" for block in response.content
)

if has_redacted_thinking:
    print("Response contains redacted thinking blocks")
    # These blocks are still usable in subsequent requests

    # Extract all blocks (both redacted and non-redacted)
    all_thinking_blocks = [
        block for block in response.content
        if block.type in ["thinking", "redacted_thinking"]
    ]

    # When passing to subsequent requests, include all blocks without modification
    # This preserves the integrity of Claude's reasoning

    print(f"Found {len(all_thinking_blocks)} thinking blocks total")
    print(f"These blocks are still billable as output tokens")

在生產環境中處理被編輯思考的建議

在建構使用延伸思考的面向客戶應用程式時:

請注意被編輯的思考區塊包含無法供人閱讀的加密內容
考慮提供簡單的說明,例如:“Claude 的部分內部推理已因安全原因自動加密。這不會影響回應品質。”
如果要向使用者顯示思考區塊,您可以過濾掉被編輯的區塊,同時保留正常的思考區塊
明確說明使用延伸思考功能可能偶爾會導致某些推理被加密
實作適當的錯誤處理,以優雅地管理被編輯的思考而不破壞您的 UI

串流延伸思考

當啟用串流時,您將通過 thinking_delta 事件接收思考內容。以下是如何處理帶有思考的串流:

curl https://api.anthropic.com/v1/messages \
     --header "x-api-key: $ANTHROPIC_API_KEY" \
     --header "anthropic-version: 2023-06-01" \
     --header "content-type: application/json" \
     --data \
'{
    "model": "claude-3-7-sonnet-20250219",
    "max_tokens": 20000,
    "stream": true,
    "thinking": {
        "type": "enabled",
        "budget_tokens": 16000
    },
    "messages": [
        {
            "role": "user",
            "content": "What is 27 * 453?"
        }
    ]
}'

串流輸出範例:

event: message_start
data: {"type": "message_start", "message": {"id": "msg_01...", "type": "message", "role": "assistant", "content": [], "model": "claude-3-7-sonnet-20250219", "stop_reason": null, "stop_sequence": null}}

event: content_block_start
data: {"type": "content_block_start", "index": 0, "content_block": {"type": "thinking", "thinking": ""}}

event: content_block_delta
data: {"type": "content_block_delta", "index": 0, "delta": {"type": "thinking_delta", "thinking": "Let me solve this step by step:\n\n1. First break down 27 * 453"}}

event: content_block_delta
data: {"type": "content_block_delta", "index": 0, "delta": {"type": "thinking_delta", "thinking": "\n2. 453 = 400 + 50 + 3"}}

// Additional thinking deltas...

event: content_block_delta
data: {"type": "content_block_delta", "index": 0, "delta": {"type": "signature_delta", "signature": "EqQBCgIYAhIM1gbcDa9GJwZA2b3hGgxBdjrkzLoky3dl1pkiMOYds..."}}

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

event: content_block_start
data: {"type": "content_block_start", "index": 1, "content_block": {"type": "text", "text": ""}}

event: content_block_delta
data: {"type": "content_block_delta", "index": 1, "delta": {"type": "text_delta", "text": "27 * 453 = 12,231"}}

// Additional text deltas...

event: content_block_stop
data: {"type": "content_block_stop", "index": 1}

event: message_delta
data: {"type": "message_delta", "delta": {"stop_reason": "end_turn", "stop_sequence": null}}

event: message_stop
data: {"type": "message_stop"}

關於思考的串流行為

當使用啟用思考的串流時,您可能會注意到文字有時會以較大的塊狀形式到達,與較小的逐字傳遞交替出現。這是預期的行為,尤其是對於思考內容。

串流系統需要批次處理內容以獲得最佳效能,這可能會導致這種”塊狀”傳遞模式。我們正在持續努力改善這種體驗,未來的更新將著重於使思考內容更流暢地串流。

redacted_thinking 區塊不會有任何增量相關,並會作為單一事件發送。

範例：帶有被編輯思考的串流

curl https://api.anthropic.com/v1/messages \
     --header "x-api-key: $ANTHROPIC_API_KEY" \
     --header "anthropic-version: 2023-06-01" \
     --header "content-type: application/json" \
     --data \
'{
    "model": "claude-3-7-sonnet-20250219",
    "max_tokens": 20000,
    "stream": true,
    "thinking": {
        "type": "enabled",
        "budget_tokens": 16000
    },
    "messages": [
        {
            "role": "user",
            "content": "ANTHROPIC_MAGIC_STRING_TRIGGER_REDACTED_THINKING_46C9A13E193C177646C7398A98432ECCCE4C1253D5E2D82641AC0E52CC2876CB"
        }
    ]
}'

這將輸出:

event: message_start
data: {"type":"message_start","message":{"id":"msg_018J5iQyrGb5Xgy5CWx3iQFB","type":"message","role":"assistant","model":"claude-3-7-sonnet-20250219","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":92,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":3}}       }

event: content_block_start
data: {"type":"content_block_start","index":0,"content_block":{"type":"redacted_thinking","data":"EvEBCoYBGAIiQAqN5Z4LumxzafxD2yf2zW+hVm/G2/Am05ChRXkU1Xe2wQLPLo0wnmoaVJI1WTkLpRYJAIz2UjzHblLwkJ59xeAqQNr5EWqMZkOr8yNcpbCO5PssXiUvEjhoaC0IN3qyhE3vumOOS9Qd0Ku4AYTgu8VjP4C6IJHnkuIexa0VrU/cFbISDJjPWOWQlyAx4y5FCRoMk55jLUCR8KCZrKrzIjDR8S3F/pCWlz/JA5RN0uprpWAI75HjgcY2NJkPX3sEC0Ew6fl6YEISNk1XsmzWtj4qGArQlCfAW9l8SDiKbXm0UZ4hQhh2ruPbaw=="}  }

event: ping
data: {"type": "ping"}

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

event: content_block_start
data: {"type":"content_block_start","index":1,"content_block":{"type":"redacted_thinking","data":"EvMBCoYBGAIiQKZ6LAz+dCNWxvz0dmjI0gfqEInA9MLVAtFJTpolzOaIbUs28xuKyXVzEQsPWPvP12gN+hxVJ4mYzWT8DCIAxXIqQHwQZcGASLMxWCfHrUlYfFq0vF8IGhRgQKxpj1zxouNLuKdhpZrcHF9vKODIPCPW8EWD13aI6t+exz/UboOs/ZMSDA8tVDp4vkOEUc7sGBoMbiRhGYMqcmAOhb3nIjC/lewBt2l9P+VpJkV78YQ3LhvNh/q3KfsbGuJ2U+lIMPGf9wnzrRC/6xqdsHPe1B0qGozBPKnbBifhyb7xYyWcEWoi/qW9OdoFl1/w"}            }

event: content_block_stop
data: {"type":"content_block_stop","index":1             }

event: content_block_start
data: {"type":"content_block_start","index":2,"content_block":{"type":"redacted_thinking","data":"Eu8BCoYBGAIiQCXtUNB4QyT//Zww832Q+xjJ0oa7/PQZr74OvbS1+a7cRNywZfYMYGGte3RXXTMa6I0bFJOMmXXckcbLxR/L+msqQLhKGx9Bt2FnLpo7bp/PdMQBDDCo+jkbOctnxBQrHCuYbu33o30qPCh73AZ8O1xXXEZfzfLC0L6RoHzLxQSHN5gSDAxGSY7Ifg073BaUYBoMSWHLVrmZrydEfc7SIjAF1R+fYlyVPFwS4Sac/Dw9caskXNF/p+Yn7RNaW9+v/jL03qsqqvemuqRGltSBfZcqFrowQipxo/ftIkEC47Ua64RzSBIe27E="}   }

event: content_block_stop
data: {"type":"content_block_stop","index":2              }

event: content_block_start
data: {"type":"content_block_start","index":3,"content_block":{"type":"redacted_thinking","data":"Eu8BCoYBGAIiQEgE6WUvQO3d6fPpY3OaA95soqeWgZv/Nyi0X6iywTb5KqvUn9NxWySiZwSFZb+4S8ymtHRO4OBKA7eRWEXcBuQqQNudvV6YSFH5ErwaDME0HaEjtHcuy8SslL6RhLwhEJKGpYCzq7zWupcMBB1g57sR8vh/JwGjr7D9sfX9jmM7EsESDEatCbzVVczyZ0TERRoMenFOToj2qn0Xmh1LIjA1WgxaMqiHhb5T4k/++UCKNMH2SEseLzTlR7uIz20qZUXDWtoVck6wc+x7lSWRKXQqFiLoTO1oG0I/lbPz1n2FgC3MH7683FU="}     }

// Additional events...

event: content_block_start
data: {"type":"content_block_start","index":58,"content_block":{"type":"redacted_thinking","data":"EuoBCoYBGAIiQJ/SxkPAgqxhKok29YrpJHRUJ0OT8ahCHKAwyhmRuUhtdmDX9+mn4gDzKNv3fVpQdB01zEPMzNY3QuTCd+1bdtEqQK6JuKHqdndbwpr81oVWb4wxd1GqF/7Jkw74IlQa27oobX+KuRkopr9Dllt/RDe7Se0sI1IkU7tJIAQCoP46OAwSDF51P09q67xhHlQ3ihoM2aOVlkghq/X0w8NlIjBMNvXYNbjhyrOcIg6kPFn2ed/KK7Cm5prYAtXCwkb4Wr5tUSoSHu9T5hKdJRbr6WsqEc7Lle7FULqMLZGkhqXyc3BA"}        }

event: content_block_stop
data: {"type":"content_block_stop","index":58            }

event: content_block_start
data: {"type":"content_block_start","index":59,"content_block":{"type":"text","text":""}              }

event: content_block_delta
data: {"type":"content_block_delta","index":59,"delta":{"type":"text_delta","text":"I'm"}        }

event: content_block_delta
data: {"type":"content_block_delta","index":59,"delta":{"type":"text_delta","text":" not"}     }

event: content_block_delta
data: {"type":"content_block_delta","index":59,"delta":{"type":"text_delta","text":" sure"}    }

// Additional text deltas...

event: content_block_delta
data: {"type":"content_block_delta","index":59,"delta":{"type":"text_delta","text":" me know what you'"}          }

event: content_block_delta
data: {"type":"content_block_delta","index":59,"delta":{"type":"text_delta","text":"d like assistance with."}   }

event: content_block_stop
data: {"type":"content_block_stop","index":59               }

event: message_delta
data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":184}  }

event: message_stop
data: {"type":"message_stop"          }

使用延伸思考時的重要考慮事項

使用思考預算: 最小預算是 1,024 個 token。我們建議從最小值開始,逐步增加思考預算,以找到適合您使用情境的最佳範圍。較高的 token 數量可能讓您獲得更全面和細緻的推理,但根據任務的不同,可能會有遞減的回報。

思考預算是一個目標而不是嚴格的限制 - 實際的 token 使用量可能會根據任務而有所不同。
由於需要額外的處理時間用於推理過程,請準備可能較長的回應時間。
當 max_tokens 大於 21,333 時,需要使用串流。

對於超過 32K 的思考預算: 我們建議對於思考預算設定在 32K 以上的工作負載使用批次處理,以避免網路問題。推動模型在 32K token 以上進行思考的請求會導致長時間運行的請求,可能會遇到系統超時和開放連接限制。

思考與其他功能的相容性:

思考與 temperature、top_p 或 top_k 修改以及強制工具使用不相容。
當啟用思考時,您無法預填回應。
對思考預算的更改會使包含訊息的快取提示前綴失效。然而,當思考參數改變時,快取的系統提示和工具定義將繼續有效。

延伸思考的定價和 token 使用

延伸思考 token 計入上下文視窗並作為輸出 token 計費。由於思考 token 被視為正常的輸出 token,它們也計入您的速率限制。在規劃 API 使用時,請務必考慮這種增加的 token 使用量。

Model	Base Input Tokens	Cache Writes	Cache Hits	Output Tokens
Claude Opus 3.8	$15 / MTok	$18.75 / MTok	$1.50 / MTok	$75 / MTok
Claude Sonnet 3.8	$3 / MTok	$3.75 / MTok	$0.30 / MTok	$15 / MTok
Claude 3.7 Sonnet	$3 / MTok	$3.75 / MTok	$0.30 / MTok	$15 / MTok

延伸思考的批次處理享有這些價格的 50% 折扣,且通常在不到 1 小時內完成。

所有延伸思考 token(包括被編輯的思考 token)都作為輸出 token 計費,並計入您的速率限制。

在多輪對話中,與較早的助理訊息相關的思考區塊不會作為輸入 token 計費。

當啟用延伸思考時,會自動包含一個專門的 28 或 29 個 token 的系統提示以支援此功能。

範例：未來回合中省略先前的思考 token 作為輸入 token

此範例展示即使第二條訊息包含助理的完整回應和思考區塊,token 計數 API 顯示先前的思考 token 不會計入後續回合的輸入 token 計數:

import anthropic

client = anthropic.Anthropic()

# First message with extended thinking enabled
first_message = [{
    "role": "user",
    "content": "Explain quantum entanglement"
}]

# Get the first response with extended thinking
response = client.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=20000,
    thinking={
        "type": "enabled",
        "budget_tokens": 16000
    },
    messages=first_message
)

# Count tokens for the first exchange (just the user input)
first_count = client.messages.count_tokens(
    model="claude-3-7-sonnet-20250219",
    messages=first_message
)
print(f"First message input tokens: {first_count.input_tokens}")

# Prepare the second exchange with the previous response (including thinking blocks)
second_message = first_message + [
    {"role": "assistant", "content": response.content},
    {"role": "user", "content": "How does this relate to quantum computing?"}
]

# Count tokens for the second exchange
second_count = client.messages.count_tokens(
    model="claude-3-7-sonnet-20250219",
    messages=second_message
)
print(f"Second message input tokens: {second_count.input_tokens}")

# Extract text-only blocks to compare
text_only_blocks = [block for block in response.content if block.type == "text"]
text_only_content = [{"type": "text", "text": block.text} for block in text_only_blocks]

# Create a message with just the text blocks for comparison
text_only_message = first_message + [
    {"role": "assistant", "content": text_only_content},
    {"role": "user", "content": "How does this relate to quantum computing?"}
]

# Count tokens for this text-only message
text_only_count = client.messages.count_tokens(
    model="claude-3-7-sonnet-20250219",
    messages=text_only_message
)

# Compare token counts to prove previous thinking blocks aren't counted
print(f"Are they equal? {second_count.input_tokens == text_only_count.input_tokens}")

這是腳本的輸出(您可能會看到略有不同的數字)

First message input tokens: { cache_creation_input_tokens: 1365, cache_read_input_tokens: 0, input_tokens: 44, output_tokens: 725 }

Second message input tokens: { cache_creation_input_tokens: 0, cache_read_input_tokens: 1365, input_tokens: 386, output_tokens: 765 }

Third message input tokens: { cache_creation_input_tokens: 0, cache_read_input_tokens: 1365, input_tokens: 811, output_tokens: 542 }

此行為發生是因為 Anthropic API 在計算上下文使用量時會自動刪除先前回合的思考區塊。這有助於優化 token 使用,同時保持延伸思考的好處。

延伸輸出功能(測試版)

Claude 3.7 Sonnet 可以產生比先前模型長得多的回應,支援最多 128K 輸出 token(測試版)—比其他 Claude 模型長 15 倍以上。這種擴展的功能對於涉及複雜推理、豐富的程式碼生成和全面內容創建的延伸思考用例特別有效。

此功能可以通過傳遞 anthropic-beta 標頭值 output-128k-2025-02-19 來啟用。

curl https://api.anthropic.com/v1/messages \
     --header "x-api-key: $ANTHROPIC_API_KEY" \
     --header "anthropic-version: 2023-06-01" \
     --header "anthropic-beta: output-128k-2025-02-19" \
     --header "content-type: application/json" \
     --data \
'{
    "model": "claude-3-7-sonnet-20250219",
    "max_tokens": 128000,
    "thinking": {
        "type": "enabled",
        "budget_tokens": 32000
    },
    "messages": [
        {
            "role": "user",
            "content": "Generate a comprehensive analysis of..."
        }
    ],
    "stream": true
}'

當使用較長輸出的延伸思考時,您可以分配較大的思考預算以支援更全面的推理,同時仍有充足的 token 用於最終回應。

我們建議對這種延伸輸出功能使用串流或批次模式;有關更多詳細資訊,請參閱我們關於長請求的網路可靠性考慮指南。

使用延伸思考與提示快取

提示快取與思考有幾個重要的考慮事項:

快取提示中包含思考區塊

思考只在生成助理回合時包含,不適合快取。
忽略先前回合的思考區塊。
如果思考被停用,傳遞給 API 的任何思考內容都會被簡單地忽略。

快取失效規則

對思考參數的更改(啟用/停用或預算變更)會使在訊息中設置的快取斷點失效。
當思考參數改變時,系統提示和工具仍然保持快取。

延伸思考與提示快取的範例

系統提示快取(當思考改變時保留)

from anthropic import Anthropic
import requests
from bs4 import BeautifulSoup

client = anthropic.Anthropic()

def fetch_article_content(url):
    response = requests.get(url)
    soup = BeautifulSoup(response.content, 'html.parser')

    # Remove script and style elements
    for script in soup(["script", "style"]):
        script.decompose()

    # Get text
    text = soup.get_text()

    # Break into lines and remove leading and trailing space on each
    lines = (line.strip() for line in text.splitlines())
    # Break multi-headlines into a line each
    chunks = (phrase.strip() for line in lines for phrase in line.split("  "))
    # Drop blank lines
    text = '\n'.join(chunk for chunk in chunks if chunk)

    return text

# Fetch the content of the article
book_url = "https://www.gutenberg.org/cache/epub/1342/pg1342.txt"
book_content = fetch_article_content(book_url)
# Use just enough text for caching (first few chapters)
LARGE_TEXT = book_content[:5000]

SYSTEM_PROMPT=[
    {
        "type": "text",
        "text": "You are an AI assistant that is tasked with literary analysis. Analyze the following text carefully.",
    },
    {
        "type": "text",
        "text": LARGE_TEXT,
        "cache_control": {"type": "ephemeral"}
    }
]

MESSAGES = [
    {
        "role": "user",
        "content": "Analyze the tone of this passage."
    }
]

# First request - establish cache
print("First request - establishing cache")
response1 = client.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=20000,
    thinking={
        "type": "enabled",
        "budget_tokens": 4000
    },
    system=SYSTEM_PROMPT,
    messages=MESSAGES
)

print(f"First response usage: {response1.usage}")

MESSAGES.append({
    "role": "assistant",
    "content": response1.content
})
MESSAGES.append({
    "role": "user",
    "content": "Analyze the characters in this passage."
})
# Second request - same thinking parameters (cache hit expected)
print("\nSecond request - same thinking parameters (cache hit expected)")
response2 = client.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=20000,
    thinking={
        "type": "enabled",
        "budget_tokens": 4000  # Same thinking budget
    },
    system=SYSTEM_PROMPT,
    messages=MESSAGES
)

print(f"Second response usage: {response2.usage}")

MESSAGES.append({
    "role": "assistant",
    "content": response2.content
})
MESSAGES.append({
    "role": "user",
    "content": "Analyze the setting in this passage."
})

# Third request - different thinking budget (cache hit expected because system prompt caching)
print("\nThird request - different thinking budget (cache hit expected)")
response3 = client.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=20000,
    thinking={
        "type": "enabled",
        "budget_tokens": 8000  # Different thinking budget - STILL maintains cache!
    },
    system=SYSTEM_PROMPT,
    messages=MESSAGES
)

print(f"Third response usage: {response3.usage}")

初步步驟

模型與定價

探索功能

瞭解 Claude

測試和評估

法律中心

代理元件

使用延伸思考來建構

延伸思考的運作方式

實作延伸思考

理解思考區塊

在生產環境中處理被編輯思考的建議

串流延伸思考

使用延伸思考時的重要考慮事項

延伸思考的定價和 token 使用

延伸輸出功能(測試版)

使用延伸思考與提示快取

延伸思考與提示快取的範例

初步步驟

模型與定價

探索功能

瞭解 Claude

測試和評估

法律中心

代理元件

​延伸思考的運作方式

​實作延伸思考

​理解思考區塊

​在生產環境中處理被編輯思考的建議

​串流延伸思考

​使用延伸思考時的重要考慮事項

​延伸思考的定價和 token 使用

​延伸輸出功能(測試版)

​使用延伸思考與提示快取

​延伸思考與提示快取的範例

延伸思考的運作方式

實作延伸思考

理解思考區塊

在生產環境中處理被編輯思考的建議

串流延伸思考

使用延伸思考時的重要考慮事項

延伸思考的定價和 token 使用

延伸輸出功能(測試版)

使用延伸思考與提示快取

延伸思考與提示快取的範例