Claude Opus 4.6: What's New and What It Means for AI-Powered Products

This is the change that requires attention. Extended thinking previously used budget_tokens — a hard cap on how many tokens Claude could spend reasoning. The problem: set it too low, and complex tasks got truncated reasoning. Set it too high, and you burned tokens on simple questions.

Adaptive thinking replaces this with an effort parameter that lets Claude decide how much thinking is appropriate:

# Old way — will be deprecated in future releases
response = client.messages.create(
    model="claude-opus-4-6",
    max_tokens=8000,
    thinking={
        "type": "enabled",
        "budget_tokens": 4000  # Fixed cap
    },
    messages=[...]
)

# New way — Claude decides how much thinking is needed
response = client.messages.create(
    model="claude-opus-4-6",
    max_tokens=8000,
    thinking={
        "type": "enabled",
        "effort": "medium"  # low | medium | high
    },
    messages=[...]
)

The key insight: A simple "summarize this paragraph" gets minimal thinking. A complex multi-step reasoning problem gets deep analysis. You're no longer guessing the right token budget — you're telling Claude how hard to try.

Migration urgency: budget_tokens still works on Opus 4.6, but Anthropic has explicitly said it will be retired in future model releases. If you're using extended thinking, plan your migration now.

The standard 200K context window jumps to 1M tokens in beta. To put that in perspective:

• 200K tokens ≈ a mid-sized codebase or 3-4 research papers
• 1M tokens ≈ an entire large codebase, or 15-20 research papers, or a full book

The pricing structure is important: requests stay at base pricing up to 200K. Beyond that, you pay 2x on input and 1.5x on output. A request using 500K input tokens costs effectively $10/M — still reasonable for scenarios where you genuinely need the full context.

Best use cases: Codebase-wide refactoring analysis, large document synthesis, comprehensive research review, and long-running agent conversations that accumulate context over many turns.

This is particularly relevant for agent-based applications. When a conversation approaches the context limit, Claude automatically summarizes older context to make room for new information. Think of it as intelligent memory management.

Instead of hitting a wall at your context limit and losing the conversation, the model gracefully compresses earlier turns while preserving the most important information. For tools like ClaudeArchitect — where orchestration can involve many specialist calls in a single session — this is a meaningful quality-of-life improvement.

The practical impact: Longer agent sessions without context window errors. More reliable multi-step workflows. Less need for manual conversation summarization in your application code.