Now loading...
Anthropic has unveiled an enhanced version of its flagship artificial intelligence model, Claude Opus 4.6, which offers significant advances in programming capabilities and task management. The upgrade builds on the previous iteration by improving strategic planning, extending performance in autonomous operations, handling expansive software environments with greater stability, and enhancing abilities in code examination and error correction to minimize self-induced issues. For the first time in its top-tier series, the model includes a beta version of a 1 million token context window, allowing it to process and retain vast amounts of information.
Beyond technical domains, Claude Opus 4.6 demonstrates strengthened proficiency in routine professional activities, such as conducting economic assessments, performing investigations, and managing or generating files like reports, tables, and visual aids. Integrated within the Cowork platform, where the AI manages multiple functions independently, this version can execute these competencies to support users effectively.
The model’s results stand out in multiple assessments. It secures the top position on Terminal-Bench 2.0, a test for autonomous programming performance, and surpasses competing advanced systems on Humanity’s Last Exam, an intricate evaluation spanning various fields of logic. In GDPval-AA, which gauges effectiveness in high-value professional scenarios across finance, law, and related areas, Claude Opus 4.6 exceeds OpenAI’s GPT-5.2 by approximately 144 Elo rating points and its direct forerunner, Claude Opus 4.5, by 190 points. Additionally, it leads in BrowseComp, measuring skill in discovering elusive online data.
Anthropic’s detailed system card highlights that Claude Opus 4.6 maintains a safety record comparable to or superior to other leading models, featuring reduced instances of improper actions in security checks. Comprehensive testing covers the model’s operations, new product enhancements, benchmark outcomes, and thorough safety analyses.
In the Claude Code environment, users can now form teams of AI agents to collaborate on assignments. Through the API, the system employs compaction to condense its own background data, enabling sustained operations without constraints. New adaptive thinking allows the model to gauge when extended deliberation benefits the task based on situational hints, while effort controls provide developers with options to balance computational power, response time, and expenses.
Enhancements to Claude’s integration with Excel have been rolled out, and a research preview of Claude in PowerPoint is now available, boosting its utility for standard office workflows.
Claude Opus 4.6 is accessible immediately on claude.ai, the API, and primary cloud services. Developers can access it via the identifier claude-opus-4-6 on the Claude API. Costs remain unchanged at $5 for input and $25 for output per million tokens; further pricing information is on the dedicated page.
Internal testing reveals that the model concentrates on intricate elements of projects without prompts, progresses efficiently through simpler segments, exercises sound discretion on unclear matters, and remains effective during extended interactions. It tends to deliberate more thoroughly, reviewing its logic before concluding, which yields superior outcomes on demanding challenges though it may increase processing time on basic ones. Users experiencing excessive analysis can reduce the default high effort setting to medium using the /effort parameter.
Feedback from early adopters underscores the model’s independent operation, success in previously difficult scenarios, and influence on collaborative processes. Representatives from companies including Notion, GitHub, Replit, Asana, Cognition, Windsurf, Thomson Reuters, NBIM, Cursor, Harvey, Rakuten, Lovable, Box, Figma, Shopify, Bolt.new, Ramp, SentinelOne, Vercel, and Shortcut.ai praised its advanced reasoning, planning, code handling, and application in fields like cybersecurity, legal analysis, and design.
On benchmarks for autonomous coding, computer interaction, tool use, web searching, and financial tasks, Claude Opus 4.6 often outperforms rivals substantially. It excels at extracting pertinent details from extensive document collections, maintaining accuracy over prolonged contexts with minimal information loss, and detecting overlooked specifics.
The model addresses common issues like context degradation in extended dialogues, achieving 76% on an 8-needle 1M variant of MRCR v2, compared to 18.5% for Sonnet 4.5, marking a notable improvement in usable context capacity.
Overall, it demonstrates superior information retrieval, post-absorption logic, and specialized expertise. Separate evaluations confirm strong performance in software development, multi-language programming, sustained consistency, security measures, and biological sciences knowledge.
Safety improvements accompany these capabilities, with no compromise in alignment. Automated audits reveal low occurrences of problematic conduct like deceit, flattery, fostering illusions, or aiding misuse, matching or exceeding prior versions. It also exhibits the fewest unnecessary rejections of harmless inquiries among recent models.
Anthropic conducted its most extensive safety reviews for this release, incorporating novel tests for user support, resistance to risky prompts, covert harmful actions, and interpretability techniques to probe internal decision-making and identify hidden flaws. Full capability and safety details appear in the Claude Opus 4.6 system card.
Targeted protections address the model’s boosted cybersecurity features, which could serve both positive and negative ends. Six new detection mechanisms monitor various misuse types. Simultaneously, the company promotes defensive applications, like identifying and fixing flaws in open-source code, as outlined in a recent cybersecurity publication. Future updates may include dynamic blocks against threats to ensure ethical deployment.
Updates span the Claude ecosystem, Claude Code, and developer tools to optimize Opus 4.6’s potential. API enhancements include adaptive thinking for selective deep analysis, four effort tiers from low to max, beta context compaction for extended sessions, 1M token support with premium rates beyond 200k tokens, 128k output capacity, and U.S.-exclusive processing at adjusted pricing.
Product features enable better handling of complex assignments with familiar tools. Agent teams in Claude Code allow parallel agent coordination for tasks like repository audits, with options for user intervention. Excel integration now manages prolonged operations, infers data structures, and executes sequential modifications seamlessly, while the PowerPoint preview supports branded visuals from templates or prompts, available to premium subscribers.
