Anthropic presents Claude Opus 4.6, its AI capable of reasoning for extended periods and coordinating teams of agents

Anthropic is attracting significant attention these days. If yesterday we informed you that the company announced it would not be integrating advertising into Claude, today we present its new AI model, the most intelligent and capable to date: Claude Opus 4.6. Among its main features are substantial improvements in coding and logical reasoning, a context window of 1 million tokens in beta version, and the introduction of the “agent teams” feature.

Claude Opus 4.6 is now available to users subscribed to the Pro, Max, Team, and Enterprise plans through claude.ai and other cloud platforms. Additionally, developers may now access it via the API at a cost of $5/$25 per million tokens.

Claude Opus 4.6: agent capabilities and enhanced reflection

Claude Opus 4.6 represents a significant qualitative leap compared to its previous version (Claude Opus 4.5), as it is capable of conducting a deeper reflection, providing greater attention to detail when planning, executing agent tasks over longer periods, operating more reliably on massive codebases, and detecting its own errors.

Its agent capabilities have indeed been elevated, and the company claims that Claude Opus 4.6 demonstrates a tendency to work autonomously without assistance. This makes it a collaborator capable of coordinating independent subtasks, executing tools in parallel, and identifying technical bottlenecks without constant human intervention. An example of this potential is the experiment led by Nicholas Carlini at Anthropic, in which a team of 16 Claude Opus 4.6 agents succeeded in building a C compiler from scratch in just two weeks.

Optimization for everyday work

Beyond code, Anthropic has optimized Opus 4.6 for everyday knowledge work. The model excelled in multidisciplinary reasoning assessments such as Humanity’s Last Exam, and in tasks of economic value within sectors such as legal and finance.

In addition to the release of Claude Opus 4.6, and to further enhance its role as a comprehensive autonomous assistant, Anthropic has announced a series of highly relevant new features:

• Agent Teams: You can now create “Agent Teams” in Claude Code, enabling you to have several agents specialized in different tasks available, coordinated autonomously, collaborating, and handling more complex processes.
• API “Compaction”: This feature, currently in beta, enables the summarization of context in long-running conversations when its “context window” (short-term memory) begins to approach its limit. In essence, Claude itself summarizes the most important aspects of the prior conversation to “make space,” allowing it to perform longer tasks without exceeding the context window limits.
• Adaptive Thinking: The AI detects if it has been asked a simple question (such as “What time is it?”) or a complex one (such as “Analyze this legal contract”). The model determines how many “deep reasoning” resources to allocate based on contextual clues.
• New Effort Controls: A “remote control” for developers. They can configure Claude to operate at four different levels of effort: low, medium, high (default), and maximum, depending on the depth of reasoning required, their budget, and their needs.
• Improvements in Claude for Excel: Claude’s AI can now handle complex, prolonged tasks in Excel with improved performance. Additionally, it can plan before acting, process unstructured data, infer the correct structure without guidance, and manage multi-step changes in a single pass.
• Claude for PowerPoint: Claude has been integrated into PowerPoint as a sidebar panel, enabling the AI to assist in creating visual presentations. Claude can interpret your slide layouts, fonts, and patterns to ensure strong brand consistency. This feature is available in research preview for Max, Team, and Enterprise plans.

• You may also be interested in: OpenAI introduces Frontier, the ultimate connective tissue for AI in enterprises

Feedback from early adopters

Anthropic has shared the testimonials of several partners who have already had the opportunity to work with Claude 4.6. Sarah Sachs, Director of Artificial Intelligence at Notion, stated: “Claude Opus 4.6 is the most powerful model Anthropic has released. It handles complex requests successfully by breaking them into concrete steps, executing them and delivering flawless work even when the task is ambitious. For Notion users, it feels less like a tool and more like a competent collaborator.”

Meanwhile, Fabian Hedin, co-founder of Lovable, expressed: “Claude Opus 4.6 marks an improvement in design quality. It integrates seamlessly with our design systems and is more autonomous, which is fundamental for Lovable’s values. People should be creating things that matter, not micro-managing the AI.”

In addition, Paulo Arruda, Staff Engineer at Shopify, highlighted: “Claude Opus 4.6 is the best model from Anthropic we have tested. It understands intent with minimal prompting and exceeded all expectations by exploring and generating details that I did not even know I wanted until I saw them. I felt like I was working with the model, not waiting for it.”

Safety controls

One of the primary concerns with increasing AI power is that it may become unpredictable or difficult to control. However, Anthropic assures that the improvements in Claude Opus 4.6 do not compromise safety. In its automated audits, the model demonstrated a minimal rate of misaligned behaviors, such as deception, flattery, or cooperation with misuse. “In general, it is as well aligned as its predecessor, Claude Opus 4.5, our most aligned frontier model to date.”

Furthermore, the company maintains that Claude Opus 4.6 shows the lowest rate of excessive refusals to date, preventing the AI from mistakenly refusing to answer entirely harmless queries.

To achieve this level, the model has undergone a rigorous set of safety assessments, including user welfare testing and advanced interpretability techniques to understand what occurs within its digital “brain.”

Nevertheless, the most pragmatic advancement lies in active cybersecurity defense. Given its increasing technical capabilities, Anthropic has developed six new cybersecurity probes to detect malicious usage, while simultaneously using Claude itself to identify and correct vulnerabilities in open-source software. The objective is to enable defenders to leverage AI to balance the playing field against threats, with an eye towards future real-time interventions to block any misuse attempts.

Photo: Anthropic