Anthropic Releases Claude Opus 4.6 With Agent Teams and 1M Context Window
Anthropic launches Claude Opus 4.6 featuring agent teams, adaptive thinking, 1M token context window, and state-of-the-art performance on Terminal-Bench 2.0 and Humanity's Last Exam.
Anthropic has released Claude Opus 4.6, a major update to its most capable model that introduces several features designed to push the boundaries of what AI agents can accomplish. The release includes agent teams, a system for splitting complex tasks across multiple coordinated agents, along with a massive 1M token context window, adaptive thinking with effort controls, and context compaction. The model achieves state-of-the-art results on both Terminal-Bench 2.0 and Humanity's Last Exam, two of the most demanding AI benchmarks in existence.
Agent Teams: Divide and Conquer
The headline feature of Claude Opus 4.6 is agent teams. Rather than running a single AI agent that handles every aspect of a complex task sequentially, developers can now orchestrate multiple Claude agents that work in parallel, each tackling a different part of the problem.
Think of it like a well-run engineering team. One agent might be researching the codebase to understand the architecture, another is writing unit tests, a third is implementing the core feature, and a coordinator agent is keeping everything aligned. Each agent has access to its own context and tools, and they communicate through a structured protocol that Anthropic has built into the API.
This matters because many real-world tasks are too complex for a single agent pass. A large refactoring project, for example, might require understanding dozens of files, coordinating changes across multiple services, and ensuring backward compatibility. Agent teams allow these tasks to be decomposed naturally, with each agent focusing on what it does best.
Early adopters report that agent teams can reduce the time to complete complex coding tasks by 40-60% compared to a single-agent approach, though results vary significantly depending on the task and how well the team is configured.
Adaptive Thinking and Effort Controls
Claude Opus 4.6 introduces a nuanced approach to reasoning through its adaptive thinking system. Developers can now specify an effort level (low, medium, high, or max) that controls how deeply the model thinks before responding.
This is more useful than it might sound. Not every query needs the model to spend time on deep reasoning. A simple formatting task or straightforward code generation can be handled at low effort, saving both time and cost. But when you are debugging a subtle concurrency issue or reasoning about a complex architectural decision, you want the model to take its time and think carefully.
At the max effort level, Claude Opus 4.6 engages in extended reasoning chains that can span thousands of tokens of internal deliberation. Anthropic reports that this mode significantly improves performance on tasks that require multi-step logical reasoning, mathematical proof construction, and complex code analysis.
The effort controls also have practical implications for cost management. Since Anthropic charges by token, being able to dial down the thinking effort for routine tasks can meaningfully reduce API costs for applications that handle a mix of simple and complex queries.
1M Token Context and Context Compaction
The expansion to a 1M token context window is a substantial upgrade from the previous 200K limit. One million tokens is roughly equivalent to 750,000 words, enough to hold an entire large codebase, a full book, or months of conversation history in a single context.
But raw context length is only useful if the model can actually make effective use of it, and this is where context compaction comes in. As conversations or agent sessions grow long, earlier portions of the context become less relevant. Context compaction is Anthropic's system for intelligently summarizing and compressing older context, preserving the essential information while freeing up space for new content.
In practice, this means that Claude Opus 4.6 can maintain coherent, productive sessions that run for hours or even days without losing track of important details. For agent workflows that involve iterating on a large codebase, this is transformative. The model no longer needs to be periodically reminded of decisions made earlier in the session.
Benchmark Results
Claude Opus 4.6 achieves state-of-the-art results on two particularly demanding benchmarks.
Terminal-Bench 2.0 evaluates models on realistic software engineering tasks performed in a terminal environment. Claude Opus 4.6 takes the top spot, edging out the recently released GPT-5.2-Codex. Anthropic attributes this performance to the model's improved ability to plan multi-step approaches and recover from errors during execution.
Humanity's Last Exam is a benchmark composed of extremely difficult questions across science, mathematics, philosophy, and other disciplines, questions designed to be at the frontier of human expert knowledge. Claude Opus 4.6's strong performance here suggests improvements not just in coding but in general reasoning and knowledge application.
What Developers Should Know
Claude Opus 4.6 is available immediately through the Anthropic API and through Claude Code, Anthropic's command-line tool for agentic coding. The agent teams feature requires the new multi-agent API endpoints, which are documented in Anthropic's updated developer documentation.
Pricing follows the existing token-based model, with the 1M context window available at a modest premium over standard pricing. The effort controls provide a natural mechanism for managing costs, since lower effort levels consume fewer output tokens.
For existing Claude users, the upgrade path is straightforward. The model is a drop-in replacement for Claude Opus 4.5 in most applications, with the new features available as opt-in capabilities.
Looking Ahead
Anthropic's release continues the pattern of rapid capability gains across the AI industry. The agent teams feature, in particular, suggests where the field is heading: away from single monolithic AI interactions and toward coordinated systems of specialized agents working together.
As these systems become more capable and more autonomous, questions about oversight and control become more pressing. Anthropic has published an accompanying research paper on their approach to agent safety, including monitoring systems for agent team interactions and guardrails on autonomous decision-making. It is a reminder that as the capabilities grow, so too must the frameworks for using them responsibly.