What is particularly new about the new AI model version Claude Opus 4.6 from Anthropic?

No more loss of context: This is what the new “Context Compaction” in Opus 4.6 brings

With the release of Claude Opus 4.6, Anthropic makes a significant statement in the rapidly evolving AI landscape, redefining what we can expect from a language model. This update marks far more than just an incremental performance improvement over its predecessor, Opus 4.5; it represents a fundamental shift towards truly agent-based workflows and deeper autonomous problem-solving. While previous models primarily functioned as reactive assistants in a linear dialogue, Opus 4.6 positions itself as a proactive partner for complex projects.

At the heart of this realignment lies an impressive technical scaling: A massive context window of up to 1 million tokens (in beta) and a doubled output capacity to 128,000 tokens enable the model to analyze entire code repositories or hundreds of pages of documentation in a single pass and generate comprehensive solutions without being artificially limited. But sheer size isn't everything – with features like Adaptive Thinking, the AI ​​now independently decides how much "thinking effort" (effort level) is needed for a task to maintain a balance between cost, speed, and depth of analysis.

Particularly revolutionary for developers and power users is the introduction of agent teams and context compaction. Instead of working through isolated tasks sequentially, users can now create coordinated AI teams that work in parallel on different aspects of a project, while intelligent summaries in the background prevent important information from being lost during long sessions (context rot). Opus 4.6 thus transforms the user's role from micromanager to strategic leader, efficiently managing AI resources – whether in software development, complex data analysis, or even office applications.

Related to this:

• The SaaS stock market crash: AI changes the rules of the game – What's behind the stock market crash of SaaS providers?

Overview: What Opus 4.6 means in the AI ​​landscape

Claude Opus 4.6 is the latest version of Anthropic's flagship model and is considered the most intelligent expansion of the Opus line to date. Compared to Opus 4.5, Anthropic is moving decisively from a "simple" successor to the next level: It's not just about more computing power, but a profound realignment in planning, context management, and agent-based work. Key differences include a massively expanded context window with up to 1 million tokens, a completely new type of "reflective" behavior (Adaptive Thinking), and the introduction of agent teams for parallel work. For developers, data analysts, and anyone working with large codebases, document collections, or lengthy conversation histories, Opus 4.6 is therefore less of a subtle optimization and more of a paradigm shift in how one collaborates with AI assistants.

Context window: 1 million tokens and why that's a game changer

One of the most striking features of Opus 4.6 is the support for a context window of up to 1 million tokens during the beta phase. By default, Opus still uses a 200,000-token context, but the option to expand this to 1 million is crucial for large projects. Theoretically, this equates to several hundred pages of code or multiple medium-sized codebases that can simultaneously be within the model's context. This makes it possible to analyze entire repositories, lengthy documentation, or extensive research materials in a single turn, without losing important information at the beginning of the conversation.

For practical users, this means two main things: First, Claude Opus 4.6 can handle more complex, longer-term tasks without constantly having to "flip back" because the context was too narrow. Second, the risk of "context rot"—that is, the deterioration of quality when the query approaches the edge of the context boundary—is reduced. In benchmarks such as Needle-in-a-Haystack tests with 1M contexts, Opus 4.6 shows significantly better results than previous Opus models, indicating that the embedding and retrieval of information across very long contexts is now considerably more robust.

128,000 token output: Longer answers and more space for complex thought processes

In parallel with the broader input context, Opus 4.6 has increased the maximum output token count to 128,000 per response. This doubles the previous limit of 64,000 tokens and opens up entirely new possibilities for detailed responses. In practice, this means that Claude no longer needs to be artificially split into several small sections when generating entire documents, complete code files, or lengthy, structured analyses. For developers, this means that Claude Opus 4.6 can process entire features or multiple files in a single step without the response being "truncated.".

This enhancement has a particularly positive impact on agent-based workflows. In such scenarios, the model needs not only the capacity to generate lengthy answers but also sufficient space to insert complex "thinking steps" before arriving at the final solution. This is important because many optimizations in Opus 4.6 target precisely this area: more planning steps, more self-reflection on errors, and more detailed reasoning. By significantly increasing output capacity, the combination of extended thinking and deep analysis becomes practically usable—without requiring the user to constantly experiment with shorter, truncated answers.

Adaptive Thinking: How Opus 4.6 decides for itself when to "think deeply"

A key paradigm shift in Opus 4.6 is the introduction of "Adaptive Thinking." Previous versions of Claude essentially offered a binary choice: either Extended Thinking was enabled (with a fixed budget of thinking tokens) or it remained disabled. In Opus 4.6, Anthropic replaces this fixed option with an adaptive system where the model itself determines how much "thinking effort" a task requires. This is based on setting an "effort" level from which the user can choose.

There are four effort levels: low, medium, high (default), and max. In practice, this means that for simple tasks, such as renaming files or formatting text, you can use low or medium to reduce latency and costs. As soon as you encounter more complex tasks like multi-part refactorings, architectural changes, or extensive code reviews, it's worth switching to high or max. At these levels, the model will almost always think "deeper," meaning it will go through more steps before delivering an answer. The so-called "max" level is exclusive to Opus 4.6 and allows Claude to think without fixed constraints—this is especially intended for very demanding, analytical tasks.

Contextual compression: How Opus 4.6 permanently “understands” long conversations

Another key feature in Opus 4.6 is the introduction of "Context Compaction" in the beta phase. Long, ongoing conversations or agent workflows tend to fill the context until they eventually reach a limit. In previous versions, this meant that the quality dropped or the session was terminated due to lack of space. Opus 4.6 addresses this problem proactively: When the conversation approaches a configurable threshold, the model automatically summarizes older content and replaces it with condensed summaries.

These summaries retain their relevant content, preserving important decisions, code changes, and previous discussions. The compaction process runs transparently in the background – the user typically receives a brief notification that the conversation is being "compacted," but the continuity of the discussion is maintained. This is a crucial advantage for developers who run agents for several hours: they can complete complex projects without constant restarts or manual adjustments. Compaction not only prevents immediate termination but also ensures that the model remains stable over extended periods and doesn't "dissipate," a common problem with other models.

Agent Teams: From Individual Agents to Teams of AI Developers

One of the most ambitious features in Opus 4.6 is the introduction of "Agent Teams." Previously, a single Claude Code window could act as an agent, processing tasks and returning results to the user. In Opus 4.6, Anthropic takes this a step further: it is now possible to launch multiple independent Claude Code agents that coordinate themselves and work in parallel. These Agent Teams are being introduced as a "research preview" in many integration platforms, meaning they are not yet fully available in all interfaces, but they are very mature.

The concept: One agent acts as a "team lead," dividing the main task and assigning responsibilities to team members. Each team member/agent has their own context window and can work independently, for example, one agent working on the backend logic while another works on the frontend component or testing. The agents can send messages to each other directly, coordinate progress, and even disagree if they prefer different solutions. In practice, this leads to significantly faster projects because multiple parts can be developed in parallel without the user having to constantly switch between different windows.

Agent teams in practice: What's changing for developers

In practice, Agent-Teams fundamentally changes the working model for developers. Instead of using a single window that processes several subtasks sequentially, an entire "team workflow" can now be initiated. The user describes the overall task—for example, "Create a web application with a backend, frontend, and tests"—and the team lead distributes the work among the members. Each agent can then work in their own environment, edit files, write code, and run tests, while the lead monitors progress and consolidates the results.

For users, this means significantly reduced iteration time. Instead of repeatedly breaking a task down into small parts and issuing new instructions each time, the AI ​​team can be assigned a larger task and autonomously complete small intermediate steps. Real-world tests have shown that agent teams significantly reduce the number of necessary interactions in complex projects. Furthermore, the barrier to initiating major redesigns or complete refactorings is lowered because the AI ​​teams can organize these tasks almost autonomously.

Improved coding skills and autonomy in handling large codebases

Opus 4.6 significantly improves Claude's coding capabilities. In benchmarks like SWE-Bench, the model achieves scores of around 72.5%, a massive improvement over previous versions. This category focuses on solving real-world software engineering problems based on actual GitHub issues. A score of 72.5% means that Claude Opus 4.6 delivers acceptable solutions in roughly three out of four cases—without requiring the user to rewrite the entire solution.

This improvement is reflected in several dimensions. First, planning is significantly better: Claude now analyzes larger codebases, gains a deeper understanding of the structure, and plans steps before writing any code. Second, autonomy has increased: Opus 4.6 can perform longer-running tasks in large codebases without losing context or structure. This includes not only writing code, but also testing, debugging, and refactoring across multiple files.

Another key aspect is the ability to recognize and correct its own errors. In previous versions, users often had to search for errors and then ask the AI ​​to fix the code. In Opus 4.6, the AI ​​is increasingly able to independently check for consistency, ensure tests have passed, and maintain a sound architecture. This combination of improved planning, broader context, and autonomous error correction makes Opus 4.6 a particularly powerful partner for developers working on medium to large projects.

A new dimension of digital transformation with 'Managed AI' (Artificial Intelligence) – Platform & B2B solution | Xpert Consulting - Image: Xpert.Digital

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

• The Managed AI Solution - Industrial AI Services: The Key to Competitiveness in the Services, Industry and Mechanical Engineering Sectors

New possibilities in the use of office tools and productivity applications

Anthropic has also optimized Opus 4.6 for use in traditional productivity applications. Experimental integrations are now available, allowing Claude to work directly within Excel or PowerPoint documents. In PowerPoint, for example, Claude can not only suggest content but also actively engage with a design system, adjust layouts, and structure slides. In Excel, the AI ​​can analyze complex calculations, suggest formulas, and optimize spreadsheet architectures.

For users who work extensively with Office files, this becomes an assistant that not only formulates text but also understands numbers and structures. Combined with the large context window, Opus 4.6 can analyze an entire presentation or a complex calculation model, recognize relationships, and provide targeted suggestions without requiring the user to explain everything step by step. These integrations are still partly in the research and preview phase, but they illustrate the direction of development: away from isolated assistants and toward an AI system integrated into the entire workflow.

Related to this:

• Anthropic presents Claude Opus 4.5: Better than Google? Excel, Code & Agents – PC control included

Effort-Level Management: How to Balance AI Intelligence, Costs, and Speed

The introduction of the four effort levels is a crucial point for many companies because it allows them to use AI intelligence in a targeted and scaled manner. In practice, this means that for simple, repetitive tasks, the effort can be set to low, ensuring a fast and cost-effective response. As soon as the tasks become more complex—for example, with architectural decisions, extensive code reviews, or complex analyses—the effort is switched to high or maximum.

This mechanism is particularly important because deep thinking and lengthy expenditures are directly linked to costs. The more thinking and the more tokens are consumed, the more expensive the request becomes. Fine-grained control allows a company, for example, to use a standard pipeline for simple tasks with low or medium settings and a separate, high-quality pipeline for critical AI decisions with maximum settings. This ensures that AI is used efficiently, both economically and in terms of content.

Agent teams, context compaction, and effort levels: How the features work together

The new features of Opus 4.6 are not designed in isolation, but rather build upon one another. In practice, agent teams, context compaction, and adaptive thinking work together to enable long-term, complex agent workflows. The agents work in parallel, while context compaction ensures that each team member remains "in context" even over extended periods. Simultaneously, the model determines how much cognitive resources are required for each individual request, depending on the selected effort level.

This interplay means that users can finally start complex projects without constantly worrying about technical limitations. Instead of constantly instructing the AI ​​which files to review again, or splitting the session because the context is too full, the workflow can run seamlessly. The agent teams can coordinate with each other, automatically summarize older, less relevant content, and simultaneously think more deeply about which steps make sense next.

Benchmarks and comparisons: Where Opus 4.6 stands compared to other models

Opus 4.6 consistently ranks at the top in numerous benchmarks – particularly in areas requiring longer-term reasoning, broader contexts, and complex agent behavior. In tests like Humanity's Last Exam, a multidisciplinary benchmark for complex, multi-stage problems, Opus 4.6 achieves the highest score of all known models. In Terminal-Bench 2.0, which focuses on agent-based coding in the shell, the model also delivers top results, highlighting Opus 4.6's strength in autonomous, terminal-based workflows.

The performance of Opus 4.6 is particularly evident in the area of ​​long contexts and agent and context compression features, as demonstrated by the benchmark results. Opus 4.6 achieves top scores in many agentic coding benchmarks: in Terminal-Bench 2.0 for agentic coding, the model scores approximately 65.4%, in OSWorld for agentic computer use, 72.7%, and in BrowseComp for agentic search, around 84%. This means that Opus 4.6 not only performs significantly better than Opus 4.5, but also better than most current competing models – especially in scenarios involving multi-stage, tool-based workflows.

In multidisciplinary benchmarks such as Humanity's Last Exam with Tools, Opus 4.6 achieves approximately 53.1%, in the Finance Agent task around 60.7%, and in office task benchmarks like GDPVal-AA an Elo score of approximately 1606. These results show that the model is not only optimized for pure programming tasks, but is also increasingly performing very well in complex, combined workflows – such as research, analysis, text creation, and presentation design.

Agentic functionality: Why Opus 4.6 Agentic is more "thinking"

Anthropic has explicitly positioned Opus 4.6 as agentic-optimized. This means the model is not just a good text generator, but a system capable of breaking down complex tasks into multiple steps, controlling tools, and self-assessing progress. In benchmarks like τ2-Bench, which tests tool-based planning in retail and telecommunications scenarios, Opus 4.6 achieves approximately 91.9% in the retail portion and 99.3% in the telecom portion. This is a significant leap compared to Opus 4.5 and indicates a substantial improvement in its ability to correctly call functions, plan multiple steps simultaneously, and detect errors.

At the same time, there are some areas where performance is slightly down – for example, with MCP Atlas, where Opus 4.6 lags somewhat behind Opus 4.5 and GPT-5.2. This suggests a trade-off: The optimization for continuous, long-term agent-type workloads and the more distributed agent coordination apparently means that some very specific, high-scaling tool orchestration scenarios are no longer quite as powerful as before. For most users, however, this is not a practical problem because the overall balance between coding, OS interaction, search, and office tasks clearly favors Opus 4.6.

Multi-document and multi-coding capabilities: How 1M context works in everyday life

The 1M token context is particularly noticeable in three scenarios: large codebases, lengthy documentation, and complex projects with many artifact-related files. In practice, Opus 4.6 can now keep track of an entire Python or JavaScript codebase with several hundred files simultaneously, something that was previously only possible with artificial partitioning and manual reloading. In tests with SWE-bench, the model achieves approximately 80.8% on SWE-bench Verified, which is almost on par with Opus 4.5 – despite a significantly larger context and more complex integrated workflows.

In document scenarios such as the analysis of legal texts (HS-BigLaw Bench) or scientific research (GPQA), Opus 4.6 has significantly improved the ability to maintain consistency across long, structured texts. The combination of broader contexts, context compression, and adaptive thinking makes it possible to derive suggestions from multiple chapters, recognize connections, and identify contradictions without requiring the user to repeatedly provide additional context fragments.

Safety, reliability and refusal rate: How Opus 4.6 deals with uncertainty

Anthropic emphasizes that Opus 4.6 is not only more powerful, but also safer and more reliable than its predecessor. In practice, this manifests itself, among other things, in a lower over-refusal rate—that is, the frequency with which the model rejects sensibly posed but potentially sensitive questions. This means that in many cases, users receive direct answers to complex, technical, or business-related questions without triggering the response function, even though the question is valid and descriptively worded.

At the same time, the model's so-called "thoughtfulness" is increased: It tends to communicate uncertainties openly, document additional assumptions, and adhere more closely to predefined guidelines when debunking or writing security or compliance documents. Benchmarks for legal or financial agent tasks show that this combination of higher reliability and clearer communication of uncertainty significantly increases its usefulness in professional environments.

Efficiency, costs and token economics: When is which effort level worthwhile?

Although Opus 4.6 is significantly more powerful, the token economy remains crucial for practical users. The effort levels low, medium, high, and max directly affect the number of thinking tokens and thus costs and response time. In many everyday tasks—such as writing short texts, formatting emails, or simply debugging small code snippets—a low or medium effort level is sufficient to maintain a good balance between quality and efficiency.

For complex, long-term agent-type workflows, the picture changes: Benchmarks show that using high or max settings leads to significant improvements, especially with Terminal-Bench 2.0, OSWorld, and multidisciplinary reasoning tasks. In these cases, the higher token consumption is justified because the overall project efficiency increases: The AI ​​requires less switching back and forth, fewer correction cycles, and less human intervention. For companies, this translates into a clear strategy: Standard workflows with lower effort, critical or complex projects with higher effort.

Agent teams versus individual agents: When is teamwork useful?

Agent teams aren't necessary for every application, but they offer real added value in certain scenarios. In single-agent scenarios, a Claude window operates with a limited context, few tools, and a fixed goal. Agent teams, on the other hand, consist of multiple independent agents that coordinate themselves, take on different roles, and can work in parallel. Benchmarks using Terminal-Bench 2.0 and OSWorld demonstrate that agent teams are significantly faster and more robust than single agents, especially in large, multi-stage projects.

In practice, an agent team becomes worthwhile when a task comprises several large subtasks, such as backend development, frontend implementation, testing, and documentation. Each agent can then be responsible for one of these areas, while the team lead takes on the integrating role and monitors the results. For smaller or highly focused tasks, the overhead of an agent team is often unnecessary, as a single agent with high effort can already deliver sufficient performance.

Future perspectives: How Opus 4.6 can change the use of AI agents

Opus 4.6 is less a single step than a paradigm shift in agent architecture. With agent teams, 1M context, context compaction, and adaptive thinking, it becomes possible to run complex projects continuously for hours or even days without constant user intervention. This allows companies to automate entire engineering, research, or productivity workflows, where AI agents not only handle individual tasks but also plan, execute, and control entire projects.

At the same time, the role of humans as "designers" and "monitors" becomes more pronounced. Users define goals, set effort levels, monitor agent teams, and make final decisions, while AI handles the operational work. In this sense, Opus 4.6 marks the transition from AI assistants to AI partners that collaborate in long-term, complex workflows rather than providing occasional assistance. For developers, data analysts, and knowledge workers, this represents a profound shift that not only increases productivity but also transforms how projects are organized and managed.

What is particularly new about Claude Opus 4.6 is

What's truly new about Claude Opus 4.6 isn't so much a single feature, but rather a bundle of profound improvements that together unlock a new level of AI agent capability. These include a context window supporting up to 1 million tokens, a tripling of output tokens to 128,000, adaptive thinking with multi-level effort, the introduction of agent teams for parallel AI work, context compression for long-term sessions, and significantly improved agent capabilities in coding, terminal use, research, and office tasks.

Opus 4.6 clearly differs from Opus 4.5 in that it is not only "better," but also enables a different usage pattern: long-term, automated workflows taken over by AI teams, while humans assume the role of strategist and quality control expert. For companies using agentic workflows in software, analytics, or knowledge work, this represents a significant improvement that is reflected in both benchmarks and daily projects.

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

Konrad Wolfenstein

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here or simply call me at +49 7348 4088 965. My email address is: [email protected]

I'm looking forward to our joint project.

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs

Benefit from Xpert.Digital's extensive, five-fold expertise in a comprehensive service package | R&D, XR, PR & Digital Visibility Optimization - Image: Xpert.Digital

Xpert.Digital possesses in-depth knowledge across various industries. This allows us to develop tailored strategies precisely aligned with the requirements and challenges of your specific market segment. By continuously analyzing market trends and monitoring industry developments, we can act proactively and offer innovative solutions. The combination of experience and expertise generates added value and provides our clients with a decisive competitive advantage.

More information here:

• Benefit from Xpert.Digital's 5 areas of expertise in one package – starting from just €500/month