Adaptive Thinking Fundamentals

Opus 4.5 had binary extended thinking — on or off. Opus 4.6 replaces this with adaptive thinking: a four-level system where the model dynamically calibrates how much cognitive effort to spend on each task.

The Four Effort Levels

graph LR
    A[Quick<br/>~500-2K tokens<br/>Simple tasks] --> B[Standard<br/>~2K-10K tokens<br/>Everyday work]
    B --> C[Deep<br/>~10K-50K tokens<br/>Complex problems]
    C --> D[Maximum<br/>~50K-200K tokens<br/>Hardest challenges]

Level	Thinking Tokens	Latency	Cost Multiplier	Best For
Quick	500–2K	1–3s	1×	Formatting, simple lookups, classification
Standard	2K–10K	3–10s	1.5–2×	Code generation, debugging, analysis
Deep	10K–50K	10–30s	3–5×	Architecture review, security audits
Maximum	50K–200K	30–120s	8–15×	Novel research, exhaustive analysis

How Adaptive Thinking Decides

When you set thinking: {"type": "adaptive"}, the model reads your prompt and internally classifies the task complexity:

Prompt analysis → Complexity estimation → Effort allocation → Thinking → Response

The model considers:

• Length and complexity of the input
• Specificity of the question
• Number of constraints to satisfy
• Whether the task requires multi-step reasoning
• Ambiguity level of the request

API Configuration

Fully Adaptive (Recommended Default)

response = client.messages.create(
    model="claude-opus-4-6-20260205",
    max_tokens=4096,
    thinking={"type": "adaptive"},  # Model chooses effort level
    messages=[{"role": "user", "content": prompt}]
)

Explicit Effort Level

# Force deep thinking for a security review
response = client.messages.create(
    model="claude-opus-4-6-20260205",
    max_tokens=8192,
    thinking={"type": "adaptive", "effort": "maximum"},
    messages=[{"role": "user", "content": security_review_prompt}]
)

Disable Thinking

# Quick responses where thinking adds no value
response = client.messages.create(
    model="claude-opus-4-6-20260205",
    max_tokens=256,
    thinking={"type": "none"},  # No thinking tokens
    messages=[{"role": "user", "content": "Format this JSON: ..."}]
)

Reading Thinking Output

response = client.messages.create(
    model="claude-opus-4-6-20260205",
    max_tokens=4096,
    thinking={"type": "adaptive"},
    messages=[{"role": "user", "content": "Find the bug in this function..."}]
)

for block in response.content:
    if block.type == "thinking":
        print(f"[Thinking — {len(block.thinking)} chars]")
        print(block.thinking[:500])  # Preview thinking process
    elif block.type == "text":
        print(f"\n[Response]")
        print(block.text)

When to Override the Automatic Level

The adaptive system is good but not perfect. Override when:

Situation	Override To	Reason
Security audit	maximum	Cannot afford to miss vulnerabilities
Simple formatting	none	Thinking wastes tokens on trivial tasks
Time-sensitive response	quick	Speed matters more than depth
Research question	deep or maximum	Thoroughness matters more than speed
Batch processing	standard	Consistent quality at predictable cost

In the next lesson, you will build effort-aware applications that automatically select the right level for each task.