Claude Opus 4.6 Introduces Adaptive Reasoning and Context Compaction for Long-Running Agents (opens in new tab)

Anthropic’s Claude Opus 4.6 introduces "Adaptive Thinking" and a "Compaction API" to solve context rot in long-running agents. The model supports a 1M token context window with 76% multi-needle retrieval accuracy. While leading benchmarks in agentic coding, independent tests show a 49% detection rate for binary backdoors, highlighting the gap between SOTA claims and production security.