AttributeError: ‘ThinkingConfigEnabled’ object has no attribute ‘budget_tokens’ (2026)

When you use Anthropic’s extended thinking mode in the Claude SDK and try to access config.budget_tokens, Python raises AttributeError: 'ThinkingConfigEnabled' object has no attribute 'budget_tokens'. The fix is simple once you know the reason: the SDK expects a dict literal, not a constructor call. Here is the exact fix, why it happens, and how to avoid it in production.

Why the AttributeError happens

Anthropic’s SDK ships ThinkingConfigEnabled as a TypedDict, not a full Python class with attribute access. When you write ThinkingConfigEnabled(type="enabled", budget_tokens=16000), Python creates the object but the resulting instance does not expose .budget_tokens as an attribute. It only exposes dictionary keys.

The recommended pattern in 2026 is to pass a dict literal directly. This works because the SDK does its own key validation and never needs attribute access on the config object.

The fix: use a dict literal

import anthropic

client = anthropic.Anthropic()

# WRONG (raises AttributeError when SDK accesses .budget_tokens internally)
# from anthropic.types import ThinkingConfigEnabled
# thinking = ThinkingConfigEnabled(type="enabled", budget_tokens=16000)

# CORRECT (2026 recommended pattern)
message = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=32000,
    thinking={
        "type": "enabled",
        "budget_tokens": 16000
    },
    messages=[{"role": "user", "content": "Solve this problem step by step"}]
)

for block in message.content:
    if block.type == "thinking":
        print("Reasoning:", block.thinking)
    elif block.type == "text":
        print("Answer:", block.text)

Debugging checklist

• Confirm your Anthropic SDK version: python -c "import anthropic; print(anthropic.__version__)". Extended thinking requires SDK 0.42+ from 2026.
• Confirm the model supports thinking. Only Opus 4.8+, Sonnet 5, and later models support extended thinking. Haiku 4.5 does not.
• Confirm budget_tokens is at least 1024 and less than max_tokens. The SDK returns a validation error if these constraints are violated.
• If you built a wrapper around the SDK, check that your wrapper is passing the thinking parameter as a plain dict, not a constructor call.

Real-world example: extended thinking in a math tutor

Here is a working production pattern for an AI math tutor that uses extended thinking to solve multi-step problems.

import anthropic

def solve_with_reasoning(problem: str, thinking_budget: int = 16000):
    """Solve a math problem with visible chain-of-thought reasoning."""
    client = anthropic.Anthropic()

    resp = client.messages.create(
        model="claude-opus-4-8",
        max_tokens=32000,
        thinking={"type": "enabled", "budget_tokens": thinking_budget},
        messages=[{
            "role": "user",
            "content": f"Solve step by step: {problem}"
        }]
    )

    reasoning = ""
    answer = ""
    for block in resp.content:
        if block.type == "thinking":
            reasoning += block.thinking
        elif block.type == "text":
            answer += block.text

    return {"reasoning": reasoning, "answer": answer}

result = solve_with_reasoning("If a train leaves NYC at 60mph and another leaves Chicago at 70mph...")
print("Chain of thought:\n", result["reasoning"][:500])
print("\nFinal answer:\n", result["answer"])

Common mistakes with extended thinking

• Setting budget_tokens higher than max_tokens. The SDK rejects this immediately with a validation error, not an AttributeError.
• Using a model that does not support thinking. Sending the thinking parameter to Haiku or older Sonnet models returns a 400 error.
• Trying to stream thinking blocks the same way as text blocks. Thinking blocks arrive in a separate event type when streaming.
• Assuming thinking output is user-facing. Anthropic explicitly recommends showing thinking to developers for debugging only, not to end users by default.