Claude Opus 4.6 Released: Agent Teams and 1M Context
Anthropic launches Claude Opus 4.6 with agent teams, 1M token context, and record benchmarks. Full breakdown of new features, pricing, and what it means.
In this article

Why This Matters
Opus 4.6 addresses three limitations that have constrained serious AI adoption: context window size, single-agent bottlenecks, and the overhead of managing when a model should think harder. The 1M token context window means you can feed entire codebases, full financial quarter datasets, or comprehensive document libraries without chunking. Agent teams mean complex multi-step projects don't have to funnel through one sequential process. And adaptive thinking means the model allocates compute where it matters most.
For developers and enterprise teams, this is the first Opus-class model that can realistically handle the scope of work previously requiring multiple tool chains or human coordination layers.
What's New in Opus 4.6
Agent Teams
The most significant addition. Instead of one agent working through tasks sequentially, agent teams split larger jobs across multiple agents, each owning its piece and coordinating directly with the others. Think of it as going from a single developer working a ticket to a small team dividing and conquering a project.
In practice, this means a code migration task could have one agent analyzing the existing codebase, another writing the new implementation, and a third running tests and catching regressions - all working in parallel. For enterprise workflows, one agent could handle data gathering while another runs analysis and a third drafts the report.
1M Token Context Window
Opus 4.5 had a 200,000 token context window. Opus 4.6 jumps to 1 million tokens in beta - a first for Opus-class models. That's roughly 750,000 words, or about 3,000 pages of text in a single conversation.
This unlocks use cases that were previously impractical: analyzing an entire codebase in one pass, processing a full quarter of financial filings, or maintaining context across extended multi-day projects without losing earlier details.
Adaptive Thinking
New configurable controls let you set how aggressively Opus 4.6 uses extended thinking. Four effort levels are available: low, medium, high (default), and max. The model also uses context clues to decide when deeper reasoning is needed, so simple queries get fast responses while complex problems get the full reasoning treatment.
This matters for cost and latency. Not every API call needs maximum reasoning depth. Adaptive thinking lets you balance thoroughness against speed and token usage, which is significant for high-volume production workloads.
Context Compaction
Long conversations in the API now use context compaction, which summarizes older conversational tokens to free up room in the context window. For extended back-and-forth tasks - debugging sessions, iterative document drafting, multi-step research - this means conversations can run much longer before hitting context limits.
Improved Coding Capabilities
Opus 4.6 improves on its predecessor's coding skills across the board. It plans more carefully before writing code, sustains agentic tasks for longer sessions, operates more reliably in large codebases, and has better code review and debugging skills to catch its own mistakes. The self-correction capability is particularly notable - the model is better at identifying when its own output has issues and fixing them before delivering results.
Read the original source on anthropic.com
PowerPoint Integration
In a research preview, Opus 4.6 integrates directly into Microsoft PowerPoint. The model reads existing slide layouts, fonts, and templates, then generates or edits slides while preserving design elements. This goes beyond text generation - it understands visual design context and maintains consistency with your existing brand assets.
Benchmark Performance
The numbers tell a clear story. Opus 4.6 sets new highs across multiple established benchmarks, with the most dramatic improvement on ARC AGI 2 - a test designed to measure general reasoning ability.
The ARC AGI 2 result is the standout. Going from 37.6% to 68.8% in a single generation isn't typical for benchmark improvements - it represents a near-doubling in general reasoning capability. And beating GPT-5.2 by 14.6 percentage points on this test suggests Anthropic has made fundamental advances in how the model approaches novel problems.
Security: 500+ Zero-Day Vulnerabilities Found
Before launch, Anthropic's frontier red team tested Opus 4.6 in a sandboxed environment to see how well it could find bugs in open-source code. The result: Claude found more than 500 previously unknown zero-day vulnerabilities using just its out-of-the-box capabilities. Each vulnerability was validated by either an Anthropic team member or an outside security researcher.
500+ zero-day vulnerabilities were discovered in open-source code by Opus 4.6 during pre-launch security testing. Every finding was independently validated. This demonstrates significant capability in automated security auditing.
This has real implications for software security. Organizations running open-source dependencies (which is nearly everyone) could use Opus 4.6 for automated vulnerability scanning at a depth that previously required specialized security teams. The dual-use concern is obvious - Anthropic's safety team was involved throughout the testing - but the defensive applications are substantial.
Pricing and Availability
Opus 4.6 is available today across all channels:
- claude.ai - Available now for Pro and Team subscribers
- Claude API - Model ID:
claude-opus-4-6 - Cloud platforms - Available on AWS Bedrock, Google Cloud Vertex AI, and other major providers
- Pricing - $5 per million input tokens, $25 per million output tokens (unchanged from Opus 4.5)
Keeping pricing flat while delivering this level of improvement is notable. The 1M token context window is in beta, so expect it to become generally available after the beta period.
What This Means for Developers
If you're building on the Claude API, the practical changes are significant. Agent teams open up architectures that weren't feasible before - you can now design systems where multiple Claude instances coordinate on complex tasks natively, rather than building that orchestration layer yourself. The 1M context window removes a constraint that forced many applications into complex RAG pipelines when simpler approaches would suffice.
Adaptive thinking is a cost optimization lever. Production applications processing thousands of requests can route simple queries at lower effort levels while reserving max effort for complex reasoning tasks. Combined with context compaction for long-running sessions, the total cost of ownership for Claude-based applications should improve even at the same per-token pricing.
Quick Takeaway
Claude Opus 4.6 delivers four meaningful advances: agent teams for parallel task execution, 5x larger context window, adaptive thinking for cost and latency control, and substantially improved reasoning benchmarks. The security testing results (500+ zero-days found) demonstrate a model that's genuinely more capable at understanding and analyzing code at scale.
For teams already using Claude, the upgrade is straightforward - same pricing, same API patterns, with the new model ID claude-opus-4-6. For teams evaluating Claude, the benchmark results and agent teams feature make this the strongest case yet for Anthropic's flagship model.
Get the Claude playbook in your inbox.
One weekly email for Claude and Claude Code users. Real workflows, no hype. Subscribe and we send you The Claude Power-User Cheatsheet.
— ¶ —

Luke Thompson
Luke Thompson is the founder of The Operations Guide, LLC and editor of The Claude Insider. Based in Jonesborough, Tennessee, he has spent years building AI-augmented business systems and automation workflows for operators and teams. He began working with large language models in production well before the current wave of consumer AI tools, integrating them into client workflows, content pipelines, and operational infrastructure. At The Claude Insider, he writes about Claude with the specificity of someone who uses it daily as a professional tool — not as a reviewer or commentator, but as a builder. His coverage focuses on what actually works: prompt patterns, API integration strategies, agentic workflows, and the real-world tradeoffs that practitioners face. He is not affiliated with Anthropic, PBC.
Articles are researched and drafted with AI assistance, reviewed and edited by Luke Thompson.
Know where AI can pay off in your company.
Take the free two-minute AI Readiness Assessment. See your score, the two gaps holding you back, and the next move worth making.


