Claude Opus 5.5: Thinking You Can't Steal

# What Is Preserved Thinking?

Distillation is the practice of extracting a model's capabilities at scale. Run thousands of prompts, collect the reasoning traces, and use them to train a cheaper model. Anthropic's September 2026 threat intelligence report says they have detected and disrupted this at industrial scale.

Preserved thinking makes this structurally harder. Every thinking block gets a cryptographic signature. When you send it back in a multi-turn conversation, the API checks two things:

The result: even if you intercept every thinking block, the traces are cryptographically bound to a specific conversation state. Replay them in a different context and the API rejects them. Train on them and you get signatures, not reasoning.

# What Changes in Your Code

If you are already using extended thinking with Claude, the API surface barely changes. The key difference: thinking blocks are now opaque and signed. You pass them back but you cannot read or modify them.

import anthropic

client = anthropic.Anthropic()

# Turn 1: model generates preserved thinking
response = client.messages.create(
    model="claude-opus-5-5-20261001",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[{"role": "user", "content": "Analyze this code..."}]
)

# The thinking block is opaque — you see type + signature
# You CANNOT read its content. You CAN pass it back.
thinking_block = response.content[0]  # type: "thinking"
text_block = response.content[1]      # type: "text"

# Turn 2: pass entire content back as assistant message
follow_up = client.messages.create(
    model="claude-opus-5-5-20261001",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[
        {"role": "user", "content": "Analyze this code..."},
        {"role": "assistant", "content": response.content},
        {"role": "user", "content": "Now refactor the auth module"}
    ]
)

# How Enforcement Works

# The 40% Cost Reduction

The cache read price is the real story for agentic workloads. In coding agents and tool-use workflows, cache reads dominate the token bill. $0.20 per million cached tokens means the effective cost of long multi-turn coding sessions drops dramatically.

Context window stays at 200K. GPT-6 Sol offers 1.05M at lower raw prices, but Anthropic is betting that cache read pricing matters more than headline rates for the workflows that drive actual spend. For a 50-turn coding session with a 100K token system prompt, Opus 5.5 cache reads cost $2.00 total. On Opus 5 that was $5.00.

# The Safety Numbers

# Migration Checklist: Opus 5 → 5.5