1M context is now generally available for Opus 4.6 and Sonnet 4.6 | Claude

Source
claudeanthropiccontext-windowllm-agentsclaude-codeprompt-engineeringai-coding

Summary

Anthropic has made the full 1M token context window generally available for Claude Opus 4.6 and Sonnet 4.6 at standard pricing - no long-context premium. The key practical impact: Claude Code Max/Team/Enterprise users on Opus 4.6 now get 1M context automatically, meaning far fewer forced compactions and more intact sessions. Opus 4.6 scores 78.3% on MRCR v2 (multi-round context retrieval), the highest reported among frontier models at that context length.

Key Insight

  • No pricing penalty: $5/$25 per million tokens for Opus 4.6, $3/$15 for Sonnet 4.6 - same rate at 900K tokens as at 9K. Previously long-context incurred a multiplier.
  • Media limits expanded: 600 images or PDF pages per request (up from lower limits).
  • 78.3% MRCR v2 score: This is the retrieval-in-context benchmark at 1M length. Useful calibration point - the model can find relevant details but is not perfect.
  • Fresh context > long context: The most practised HN users strongly recommend starting new sessions rather than riding compactions deep into context. CLAUDE.md instructions dilute through compactions by default unless a hook re-inserts them.
  • Subagents are the real unlock: Each subagent starts with a clean context; the orchestrator only sees results, keeping its own context low. This lets you churn through millions of output tokens across a session without degradation - documented at code.claude.com/docs/en/agent-teams.
  • Practical cap in use: Experienced users rarely need 1M raw. Workflows using code maps (Flash for indexing) + auto-context (targeted file selection) keep individual requests to 30K–80K even on large repos. 1M is a ceiling, not a target.
  • Context poisoning is real: Bad conversation history degrades performance. Rollback (Escape twice or /rewind) rather than steering is the correct recovery pattern.
  • AI-generated slop risk: Non-technical users building apps with Opus alone produce working but insecure code. Production readiness is not the same as task completion.