Claude Opus 4.6: Agent Teams, 1M Context, and Benchmarks

Mark Neal

08 Feb 2026 — 8 min read

Anthropic has released Claude Opus 4.6, marking a significant leap forward in enterprise AI capabilities. This latest iteration introduces a 1 million token context window in beta, multi-agent coordination through Agent Teams, and benchmark results that position it as the leading model for complex knowledge work. The release comes just three months after Opus 4.5, but the improvements are far from incremental.

The most striking achievement is Claude Opus 4.6's performance on the GDPval-AA benchmark, where it scored an Elo rating of 1,606 - beating GPT-5.2 by 144 points and its predecessor Opus 4.5 by 190 points. This translates to approximately a 70% win rate in head-to-head comparisons with OpenAI's flagship model on real-world enterprise tasks. From legal document review to financial analysis and large-scale code migrations, Opus 4.6 demonstrates capabilities that fundamentally change what's possible with AI-assisted work.

What makes this release particularly noteworthy is that these performance gains come without a price increase. Opus 4.6 maintains the same pricing structure as its predecessor: $5 per million input tokens and $25 per million output tokens. For enterprises already invested in the Claude ecosystem, this represents a substantial upgrade at no additional cost.

The Context Window Revolution

The expansion from 200,000 to 1 million tokens represents more than a numerical increase. Previous long-context models suffered from what researchers call "context rot" - a degradation in performance as the context window fills. Claude Opus 4.6 addresses this problem with remarkable effectiveness.

On the MRCR v2 benchmark, which tests a model's ability to find specific information within massive contexts using an 8-needle variant across 1 million tokens, Opus 4.6 achieved a 76% accuracy score. To understand the significance of this result, consider that Claude Sonnet 4.5 scored just 18.5% on the same test. This represents a fourfold improvement in retrieval accuracy, demonstrating that Opus 4.6 can actually use its expanded context window effectively rather than simply having it available in theory.

The practical implications are substantial. Development teams can now load entire codebases into a single conversation, including hundreds of files, dependency chains, and documentation. Legal teams can process comprehensive contract portfolios without splitting them across multiple sessions. Financial analysts can work with complete annual reports, regulatory filings, and market data simultaneously. The model maintains coherence and accuracy across these massive information sets, catching details that would typically require manual cross-referencing.

This capability extends to sustained agentic sessions as well. The model maintains focus across extended multi-step operations without the degradation that plagued earlier iterations. When working on complex tasks that require dozens or hundreds of sequential actions, Opus 4.6 demonstrates superior self-correction capabilities, catching its own mistakes during code review and debugging processes.

Agent Teams: Parallel AI Coordination

Perhaps the most innovative feature in Claude Opus 4.6 is Agent Teams, introduced through Claude Code. This functionality allows multiple Claude agents to work in parallel on different aspects of a single project, coordinating autonomously while maintaining alignment with the overall objective.

The architecture enables genuinely parallel workflows. One agent might write backend API endpoints while another simultaneously develops the React frontend. A third agent can conduct security reviews on the code being produced in real-time. This isn't sequential handoff between tools - it's true concurrent operation with inter-agent coordination.

The performance metrics for Agent Teams are impressive. On BrowseComp, a benchmark measuring the ability to find difficult-to-locate information online, Opus 4.6 with Agent Teams achieved an 86.8% success rate. This represents a 24% improvement over the nearest competitor, demonstrating how effectively the parallel agent architecture handles complex information retrieval tasks that require multiple search strategies and cross-referencing.

For software development workflows, Agent Teams fundamentally changes the economics of AI-assisted coding. Large codebase reviews that might take a single agent hours can be completed in minutes when distributed across multiple coordinated agents. Each agent maintains awareness of the others' work, preventing conflicts and ensuring consistency across the entire project.

Record-Breaking Benchmark Performance

Claude Opus 4.6 achieves state-of-the-art or competitive results across nearly every major AI benchmark, with particularly strong showings in agentic and enterprise-focused evaluations.

In agentic coding, Terminal-Bench 2.0 measures a model's ability to complete complex terminal-based programming tasks. Opus 4.6 scored 65.4%, narrowly edging out GPT-5.2's Codex CLI at 64.7% and substantially ahead of Gemini 3 Pro's 56.2%. On SWE-bench Verified, which tests real-world software engineering tasks, Opus 4.6 achieved 80.8%, matching the performance tier of the best models available.

The Finance Agent benchmark evaluates models on complex financial analysis tasks. Here, Opus 4.6 scored 60.7%, significantly outperforming Gemini 3 Pro at 44.1% and GPT-5.2 at 56.6%. This benchmark is particularly relevant for enterprise adoption, as it directly measures capabilities needed for actual business workflows.

One of the most dramatic improvements appears in ARC AGI 2, a benchmark designed to test novel problem-solving abilities that require genuine reasoning rather than pattern matching. Opus 4.6 scored 68.8% - nearly double the 37.6% achieved by Opus 4.5. This massive jump suggests fundamental improvements in the model's reasoning architecture.

Visual reasoning capabilities also saw substantial gains. On MMMU Pro without tools, Opus 4.6 scored 73.9%, though Gemini 3 Pro leads this category at 81.0%. With tools enabled, Opus 4.6 reached 77.3%, demonstrating strong multimodal capabilities even if not yet best-in-class.

Specialized Domain Excellence

Beyond general benchmarks, Opus 4.6 demonstrates exceptional performance in specialized professional domains. In legal applications, the model scored 90.2% on BigLaw Bench, with 40% of responses achieving perfect scores. This benchmark specifically tests the kinds of complex legal reasoning, document analysis, and precedent application that large law firms require.

Scientific domains saw remarkable improvements as well. Opus 4.6 performs twice as well as Opus 4.5 on tasks involving computational biology, structural biology, organic chemistry, and phylogenetics. These aren't marginal gains - they represent a qualitative shift in the model's ability to handle specialized scientific reasoning that requires both domain knowledge and complex analytical capabilities.

Cybersecurity represents another area of significant advancement. During pre-release testing, Opus 4.6 identified over 500 real zero-day vulnerabilities. The model outperforms all competitors in cybersecurity vulnerability detection, combining code analysis capabilities with security-specific reasoning patterns. Anthropic introduced new cybersecurity probes in response to these enhanced capabilities, ensuring the model's defensive capabilities are properly evaluated.

The model's performance on BrowseComp deserves special attention. This benchmark tests the ability to find difficult-to-locate information online - the kind of research task that often stumps both humans and AI systems. Opus 4.6 achieved an 84.0% success rate, 24 percentage points ahead of the nearest competitor. This reflects broader improvements in long-context reasoning and retrieval, addressing the "context rot" problem that has plagued other models attempting similar tasks.

Enhanced Developer Controls and API Features

Anthropic has introduced several new API features that give developers finer control over how Opus 4.6 operates. The most significant is adaptive thinking, which replaces the previous extended thinking mode with a more flexible system.

Adaptive thinking offers four effort levels: low, medium, high, and max. The model dynamically decides when deeper reasoning helps, with high as the default setting. This allows developers to balance performance against token usage and cost depending on task requirements. Simple queries can use low effort for faster, cheaper responses, while complex analytical tasks can leverage max effort for the most thorough reasoning.

The context compaction API addresses a long-standing challenge with extended conversations. As conversations approach the context limit, server-side context summarization automatically compresses older messages, effectively enabling infinite conversations without losing critical information. This feature is particularly valuable for long-running agentic workflows where maintaining conversation history is essential but managing context manually would be impractical.

Output length has also been expanded. Opus 4.6 can now generate up to 128,000 tokens in a single response, enabling it to produce comprehensive documentation, extensive code implementations, or detailed analytical reports without requiring multiple sequential generations. For enterprise users concerned about data residency, US-only inference is now available at a premium price point.

Microsoft 365 Integration and Enterprise Features

Claude Opus 4.6 integrates directly with Microsoft 365 tools, bringing AI capabilities into the workflows where enterprise users actually work. The Excel integration allows the model to process spreadsheets for business analytics, performing complex data analysis, generating visualizations, and creating financial models directly within familiar tools.

PowerPoint integration is available as a research preview for Max, Team, and Enterprise users. This enables AI-assisted presentation creation, where Opus 4.6 can analyze data, generate slide content, and structure presentations based on high-level objectives. The combination of strong analytical capabilities and visual reasoning makes this particularly effective for executive briefings and client presentations.

These integrations reflect Anthropic's enterprise focus. Rather than requiring users to adopt entirely new workflows or platforms, Claude Opus 4.6 meets them where they already work. The model can understand spreadsheet structures, interpret charts and graphs, and generate outputs formatted appropriately for business contexts.

Enterprise users report strong performance on complex real-world tasks. Code migration projects that might take weeks of manual effort can be completed in days with AI assistance. Legal document review processes that required teams of associates can be handled more efficiently while maintaining accuracy. IT issue resolution benefits from the model's ability to understand complex system architectures and troubleshoot across multiple interconnected components.

Safety, Cost, and Availability

According to Anthropic's system card, Opus 4.6 shows low rates of misaligned behavior such as deception or over-compliance. The company reports fewer unnecessary refusals compared to previous Claude models, striking a better balance between safety and usefulness. This is particularly important for enterprise deployments where overly cautious refusals can disrupt legitimate workflows.

The pricing structure remains unchanged from Opus 4.5: $5 per million input tokens and $25 per million output tokens. However, the adaptive thinking mode does increase token consumption. Testing on the GDPval-AA benchmark showed that Opus 4.6 used 30-60% more tokens than Opus 4.5, consuming approximately 160 million tokens to complete 220 tasks. Despite using more tokens, this was still significantly less than GPT-5.2 in high-effort mode.

The combination of high per-token pricing and increased token usage makes Opus 4.6 the most costly model tested on enterprise benchmarks to date. However, for tasks where quality and accuracy are paramount - legal analysis, financial modeling, critical code reviews - the superior performance may justify the additional cost. The key is matching the effort level to the task requirements, using lower effort settings for routine work and reserving maximum effort for critical analyses.

Claude Opus 4.6 is available immediately on claude.ai, via the API, and across major cloud platforms. The 1 million token context window is currently in beta, suggesting Anthropic is gathering real-world usage data before declaring it production-ready. Enterprise customers can begin deploying the model today without waiting for lengthy procurement processes.

Conclusion

Claude Opus 4.6 represents a substantial advancement in enterprise AI capabilities. The 1 million token context window with dramatically reduced context rot, Agent Teams for parallel coordination, and record-breaking benchmark performance combine to create a model genuinely suited for complex knowledge work.

The 144-point Elo advantage over GPT-5.2 on enterprise tasks isn't just a number - it translates to meaningful differences in output quality, accuracy, and reliability. When a model wins 70% of head-to-head comparisons, users notice the difference in their daily workflows. The improvements in specialized domains like legal analysis, scientific reasoning, and cybersecurity make Opus 4.6 particularly valuable for professional applications where accuracy is non-negotiable.

Agent Teams introduces a new paradigm for AI-assisted work. Rather than a single agent working sequentially through tasks, multiple coordinated agents can tackle different aspects of a problem simultaneously. This architectural shift has implications beyond immediate performance gains - it suggests new ways of structuring AI workflows that may become standard across the industry.

The pricing stability is noteworthy. Delivering these improvements without increasing costs makes adoption easier for existing Claude users and strengthens Anthropic's competitive position. While the increased token usage in adaptive thinking mode does raise effective costs, the ability to tune effort levels gives users control over the cost-performance tradeoff.

For enterprises evaluating AI solutions, Claude Opus 4.6 sets a new standard for what's possible in AI-assisted knowledge work. The combination of technical capabilities, enterprise integrations, and safety features positions it as a serious contender for mission-critical applications where reliability matters as much as raw performance.