Anthropic Introduces Claude Opus 4.6 AI Model With Improved Agentic Performance

Anthropic released the Claude Opus 4.6 artificial intelligence (AI) model on Thursday, bringing a major update to its Opus 4 foundation model. The new model addresses key limitations in the predecessor by improving sustained performance across complex tasks, particularly in software engineering and knowledge-intensive domains. The Claude Opus 4.5 model was adept at handling advanced reasoning-based tasks, but often struggled with long-horizon contexts and edge cases in large databases. A beta version of Opus 4.6 fixes that with a context window of one million tokens.

Claude Opus 4.6 AI Model Is Here

In a newsroom post, Anthropic announced and detailed the latest frontier model. Marking a first for the Opus lineup, a beta version of the model now supports a context window of up to one million tokens, allowing it to process vast amounts of information while minimising performance degradation during long interactions. This is a step up from the 2,00,000 token limits in earlier Opus models.

Claude Opus 4.6 also comes with new features like context compaction that summarise and refresh older data during prolonged tasks. The architecture incorporates adaptive thinking, where the model assesses query complexity to allocate deeper reasoning as needed, and effort controls ranging from low to max for optimising speed, intelligence and cost.

On benchmarks, Claude Opus 4.6 sets new highs. Based on internal evaluations from the company, the post claims that the AI model leads frontier models on Terminal-Bench 2.0 for command-line proficiency, and Humanity's Last Exam for multidisciplinary reasoning. In agentic evaluations like GDPval-AA, it surpasses OpenAI's GPT-5.2 by about 144 Elo points and its own Opus 4.5 by 190 points, focusing on finance and legal tasks. SWE-bench Verified scores average 81.42 percent with optimised prompting, while CyberGym tests show strong no-thinking baseline performance.

Anthropic highlights that safeguards remain a core priority with the new model. It is said to match or exceed peers in safety audits, with low rates of deception or sycophancy and the lowest over-refusal tendencies among recent releases. The company said it added six new cybersecurity probes to detect potential misuse, accelerating defensive applications like vulnerability hunting in open-source code.

In coding, Opus 4.6 manages large repositories autonomously, conducts code reviews and debugs with high accuracy. It also assembles agent teams for parallel development via Claude Code's research preview. For business workflows, it runs financial analyses, generates documents and handles multi-step searches in tools like Claude in Excel, now upgraded for unstructured data and long-running tasks. A research preview for Claude in PowerPoint extends this to presentations. In domains like computational biology, it delivers nearly double the performance of Opus 4.5, aiding scientific discovery.

Claude Opus 4.6 is available now via the website, mobile and desktop apps, Anthropic's application programming interface (API), and major cloud providers. API pricing starts at $5 (roughly Rs. 453) per million input tokens and $25 (roughly Rs. 2,300) for output, with premiums for extended contexts.