Anthropic has unveiled Claude Opus 4.6, a significant upgrade to its flagship model designed to handle complex, long-horizon tasks with unprecedented autonomy. The model introduces a massive 1-million-token context window—the amount of information the AI can process in one go—allowing it to analyze entire codebases or dense legal libraries without losing track of details. To maintain performance across these vast datasets, Anthropic implemented "context compaction," which automatically summarizes older parts of a conversation to prevent the model from hitting its memory limits.

Beyond raw capacity, Opus 4.6 features "adaptive thinking," a capability where the model autonomously decides when a problem requires deep, step-by-step reasoning versus a quick, direct response. This is paired with new "effort controls," giving users a dial to choose between maximum intelligence for high-stakes debugging and lower-latency modes for routine tasks. In real-world benchmarks, the model outperformed competitors like GPT-5.2 in economically valuable knowledge work, particularly in finance and legal domains, while setting new records for agentic coding.

Safety remains a central pillar of the release, with the integration of six new cybersecurity probes designed to detect and block malicious use cases. The company is also utilizing the model’s own capabilities to patch vulnerabilities in open-source software, aiming to flip the script on AI-driven cyber threats. By combining sophisticated interpretability methods with robust automated audits, Anthropic claims Opus 4.6 maintains the highest standards of alignment while significantly reducing "over-refusals," where the AI mistakenly declines to answer harmless questions.

Anthropic has unveiled Claude Opus 4.6, a significant upgrade to its flagship model designed to handle complex, long-horizon tasks with unprecedented autonomy. The model introduces a massive 1-million-token context window—the amount of information the AI can process in one go—allowing it to analyze entire codebases or dense legal libraries without losing track of details. To maintain performance across these vast datasets, Anthropic implemented "context compaction," which automatically summarizes older parts of a conversation to prevent the model from hitting its memory limits.

Beyond raw capacity, Opus 4.6 features "adaptive thinking," a capability where the model autonomously decides when a problem requires deep, step-by-step reasoning versus a quick, direct response. This is paired with new "effort controls," giving users a dial to choose between maximum intelligence for high-stakes debugging and lower-latency modes for routine tasks. In real-world benchmarks, the model outperformed competitors like GPT-5.2 in economically valuable knowledge work, particularly in finance and legal domains, while setting new records for agentic coding.

Safety remains a central pillar of the release, with the integration of six new cybersecurity probes designed to detect and block malicious use cases. The company is also utilizing the model’s own capabilities to patch vulnerabilities in open-source software, aiming to flip the script on AI-driven cyber threats. By combining sophisticated interpretability methods with robust automated audits, Anthropic claims Opus 4.6 maintains the highest standards of alignment while significantly reducing "over-refusals," where the AI mistakenly declines to answer harmless questions.