Think Mode Context Compaction

Think Mode with Claude Opus 4.6 enables indefinite conversations through automatic context compaction. This server-side feature summarizes older messages when conversations approach token limits, preserving essential context while freeing space for continued analysis—no user action required.

Why Context Compaction Matters

Complex compliance work often requires extended back-and-forth analysis: comprehensive gap assessments spanning multiple controls, iterative policy refinement across dozens of exchanges, or multi-framework mappings that build on previous decisions. Traditional AI conversations hit hard token limits—typically around 200,000 tokens—forcing you to restart and lose accumulated context.

Think Mode's automatic compaction eliminates this barrier. When your conversation approaches ~150,000 tokens (roughly 75-100 message exchanges depending on length and file uploads), the system automatically summarizes earlier messages in the background. You continue working seamlessly, without losing key compliance context like control requirements discussed, policy decisions made, or risk assessments completed.

How Automated Summarization Works

Compaction is fully automatic and server-side. You don't configure anything, toggle any settings, or manually trigger the process. The system monitors conversation length continuously.

What happens:

1. Detection: When your conversation reaches ~150,000 tokens, the backend triggers compaction automatically.

2. Pause: Message streaming pauses briefly. You'll see an animated indicator: "Compacting our conversation so we can keep chatting…" with a pulsing star icon and progress bar.

3. Summarization: The system condenses older messages while preserving critical context—control references, policy frameworks, risk findings, and implementation decisions remain accessible.

4. Resume: After a few seconds, the indicator disappears and your response streams normally. The conversation continues with full effective context.

Your most recent exchanges always remain unchanged. Older messages are summarized intelligently, retaining compliance-specific details that inform ongoing work.

Claude Opus 4.6 Benefits for Long-Form Analysis

Think Mode uses Claude Opus 4.6, Anthropic's most capable reasoning model. Combined with automatic compaction, this enables uniquely powerful long-form compliance workflows:

Extended gap analysis. Upload multiple policy documents sequentially, analyze each against ISO 27001 or SOC 2 controls, build a cumulative findings list, and iterate on remediation priorities—all in one conversation. The AI remembers earlier document assessments and cross-references them as you progress.

Iterative policy development. Draft an access control policy, refine it through 15-20 rounds of revisions based on your organization's specifics, then move to incident response policy development in the same thread. Opus 4.6's deep reasoning ensures each iteration builds coherently on previous decisions.

Multi-framework mappings. Map ISO 27001 Annex A controls to SOC 2 Trust Service Criteria, add GDPR Article 32 requirements, overlay NIS2 obligations, and generate cross-framework compliance matrices—without restarting or losing cross-references between frameworks.

Complex risk assessments. Model multiple threat scenarios, evaluate likelihood and impact for each, refine assumptions based on organizational context provided incrementally, and build comprehensive risk treatment plans. Opus 4.6 maintains logical consistency across dozens of interconnected risk decisions.

Server-Side Processing: Zero Customer Action

Context compaction is completely transparent. You never configure, enable, or manually trigger it. The process runs entirely server-side:

• No settings to adjust. Think Mode includes compaction by default. No account settings, workspace preferences, or per-conversation toggles exist.

• No usage quotas affected. Compaction itself doesn't consume additional message credits. Your plan's usage limits apply to actual messages sent, not background summarization.

• No data loss. Summarization is intelligent and context-aware. Key compliance details—control numbers, policy clauses, risk ratings, implementation deadlines—are preserved in compressed form.

• No manual restarts needed. Unlike Fast mode, where you must start new conversations after 15-20 messages to avoid performance degradation, Think Mode conversations can continue indefinitely.

The only user-visible element is the brief compaction indicator. Once it completes, you continue as if nothing happened—except your conversation can now absorb another ~150,000 tokens before the next compaction cycle.

Tips for High-Quality Outputs in Long Conversations

While compaction preserves context automatically, you can optimize long Think Mode conversations for maximum accuracy and coherence:

Front-load organizational context. In your first few messages, provide specific details about your organization's size, industry, existing controls, and compliance scope. This foundational context gets summarized efficiently and informs all subsequent analysis.

Reference key decisions explicitly. When building on earlier work, briefly restate critical decisions: "Based on the risk assessment where we rated phishing as High likelihood, now let's design MFA implementation for A.5.17…" This reinforces context even after compaction.

Use structured requests. Ask for numbered lists, tables, or bulleted outputs. Structured formats make it easier for Opus 4.6 to maintain consistency across compaction cycles and for you to reference earlier findings.

Upload files at conversation start. Add policy documents, control matrices, or audit reports in your first 5-10 messages. File content consumes significant tokens; uploading early ensures summarization optimizes around your core documents rather than conversational back-and-forth.

Periodically summarize progress. Every 30-40 messages, ask the AI to recap key findings or decisions. Example: "Summarize the 8 ISO 27001 gaps we've identified so far and their priority levels." This creates explicit checkpoints that survive compaction intact.

Avoid topic drift. Even with indefinite length, focused conversations yield better results. If you finish gap analysis and want to start vendor risk assessment, consider starting a new conversation. Compaction preserves context, but clarity comes from topical coherence.

When to Use Think Mode vs. Fast Mode

Choosing the right mode optimizes both output quality and usage efficiency:

Use Think Mode (with compaction) for:

• Gap analyses requiring iterative document review and cumulative findings

• Policy development with 15+ rounds of refinement

• Multi-framework compliance mappings spanning 50+ controls

• Risk assessments evaluating interconnected threat scenarios

• Audit preparation building comprehensive evidence documentation

Use Fast Mode for:

• Quick control lookups ("What does ISO 27001 A.8.1 require?")

• Single policy template generation without extensive customization

• Isolated questions not building on previous analysis

• Tasks completable in under 10 message exchanges

You can switch modes mid-conversation, but context doesn't transfer. If you start in Fast mode and hit length limits, switching to Think mode requires manually summarizing key context from the Fast conversation in your first Think message.

Recognizing When Compaction Occurs

The compaction indicator is your only signal:

• Visual: Amber-colored banner at the top of the chat, replacing the standard typing indicator

• Icon: Pulsing star animation

• Text: Rotating messages like "Compacting our conversation so we can keep chatting…" or "Summarizing earlier messages…"

• Progress bar: Animated bar stopping at 95%, then completing when done

• Duration: 3-8 seconds on average

After compaction completes, the indicator disappears and your AI response begins streaming immediately. No action required on your part.

Plan Availability

Think Mode requires a paid subscription:

• Free plan: No access to Think Mode. Fast mode only, with 300,000 token per-conversation limit.

• Plus plan ($24/month or $240/year): Full Think Mode access with automatic compaction.

• Pro ($100/month): Think Mode included for teams and enterprises.

All Think Mode conversations benefit from compaction automatically—no separate feature toggle or premium tier required.

Related Resources

• Thinking Mode — Choosing between Fast and Think modes

• Managing Long Conversations and Usage — Conversation strategies and credit optimization

• Conversation Too Long Error — Understanding token limits and what to do when Fast mode hits length limits

• Subscription Plans and Pricing — Plan comparison and Think Mode access