Claude Opus 4.6 Thinking

Claude Opus 4.6 is designed for high-stakes professional work, positioning itself as a capable partner for complex, sustained tasks that require both depth and endurance. The model introduces adaptive thinking, a dynamic reasoning approach that lets it self-select among four effort levels—low, medium, high, and max—to allocate computational attention where it matters most. This design philosophy extends to its context handling: a 1 million token context window enters beta with Opus 4.6, achieving 76% on the MRCR v2 8-needle benchmark at that scale compared to just 18.5% for its predecessor, representing a qualitative leap in long-document reliability. The Compaction API further complements this by automatically compressing older messages as conversations approach context limits, enabling effectively infinite dialogue without manual management. Anthropic built this model to handle sophisticated professional tasks, and its ability to sustain longer work sessions and produce high-quality output reflects that intent.

The model draws from Anthropic's focus on AI safety and steerability as a research-driven organization, and its architecture is engineered for agentic workflows—tasks where the model operates autonomously across multiple steps. Opus 4.6 achieved 65.4% on Terminal-Bench 2.0, 80.8% on SWE-bench Verified, and 72.7% on OSWorld, setting the highest agentic coding scores Anthropic had recorded at launch. These benchmarks reflect genuine progress in real-world task completion rather than synthetic evaluations, which aligns with the company's stated goal of building reliable and interpretable systems. The subsequent release of Claude Opus 4.7 notably described a step-change improvement in agentic coding over Opus 4.6, indicating that the earlier model laid important groundwork. For developers integrating via API, the model supports structured outputs and system prompt instructions while disabling assistant message prefilling, signaling a deliberate shift toward more controllable and predictable agent behavior in production environments.