Anthropic Unveils Claude Opus 4.6 With a 1M-Token Context Window

• Model release and positioning

  In a blog post published Thursday, Anthropic introduced Claude Opus 4.6, its smartest model and a step up from Claude Opus 4.5 (released November 2025). CNBC calls it Anthropic’s first major model launch of 2026, following late-2025 releases of Claude Sonnet 4.5 (September 2025) and Claude Haiku 4.5 (October 2025). Opus 4.6 targets better performance in large codebases, stronger code review and debugging, and longer agentic sessions, building on the 2025 “vibe-coding” wave tied to Claude Code.

• Core technical improvements

  Opus 4.6 is designed to break tasks into steps, plan before acting, and re-check reasoning, sometimes making multiple attempts without extra prompts. Anthropic says it focuses on the hardest parts automatically, moves faster through simpler steps, handles ambiguity better, and stays productive over longer sessions.

• Effort-level control to manage overthinking

  Deeper reasoning can improve hard results but add cost and latency on simpler tasks. Anthropic recommends lowering effort from default high to medium when it overthinks. The platform now offers four effort levels: low, medium, high (default), max, plus adaptive thinking, letting the model decide when extended reasoning is needed.

• Market reaction and software industry impact

  After the launch of Claude Opus 4.6, there was a concrete signal in the market: the WisdomTree Cloud Computing Fund is down more than 20% year-to-date. It seems that the narrative is shifting from “vibe coding” to “vibe working.” Cowork-style workflows are part of broader market jitters, alongside sensitivity seen in prior model-driven moves such as DeepSeek-related shocks.

• Availability, pricing, and deployment details

  Opus 4.6 is available on Claude.ai, the Claude API (claude-opus-4-6), and “all major cloud platforms,” per Anthropic and CNBC. API pricing remains $5 per million input tokens and $25 per million output tokens. For prompts above 200k tokens, pricing rises to $10 per million input and $37.50 per million output. US-only inference is available at 1.1× token pricing.

  Opus 4.6 is on Pro/Max/Team/Enterprise; Pro costs $20/month (or $17/month annually) and can hit usage limits after long sessions.

• Long-context and long-running agent upgrades

  Opus 4.6 adds a 1M-token context window (beta)—a first for Opus—and supports up to 128k output tokens, enabling much larger single-pass work. Anthropic also introduced context compaction (beta) to summarize older conversation content and prevent agents from hitting limits. The 1M context level aligns Opus with Sonnet’s context capacity in that report’s framing, helping with larger codebases and documents. Anthropic also says Opus 4.6 reduces “context rot,” improves long-context retrieval, and finds buried details better than Opus 4.5.

• Benchmarks and comparative performance claims

  Anthropic says Opus 4.6 leads on multiple evaluations: top score on Terminal-Bench 2.0, leadership on Humanity’s Last Exam, strong BrowseComp results, and on GDPval-AA it beats OpenAI GPT-5.2 by about 144 Elo and improves over Opus 4.5 by 190 points. For long-context retrieval, Anthropic reports 76% on the “8-needle 1M” MRCR v2 test versus 18.5% for Sonnet 4.5. CNBC adds that Opus 4.6 holds the top spot on the Finance Agent benchmark.

• Product ecosystem updates for teams and office workflows

  Opus 4.6 strengthens Cowork for multitasking across research, finance, and office files (docs, spreadsheets, presentations). Claude Code adds agent teams (research preview), splitting big tasks across multiple parallel agents; users can take over subagents via keyboard navigation or tmux. Claude now works directly inside PowerPoint as a side panel, moving beyond “generate then transfer” workflows. Anthropic also cites upgrades to Claude in Excel and a PowerPoint research preview for Max/Team/Enterprise.

• Safety, alignment, and cybersecurity posture

  Anthropic says safety improved alongside capability: low rates of deception, sycophancy, user-delusion encouragement, and misuse cooperation, plus fewer “over-refusals.” It also says it built six new cybersecurity probes and is pushing defensive use cases, including finding and patching open-source vulnerabilities.

Why This Matters

Claude Opus 4.6 pushes AI beyond faster coding into “vibe working,” where systems can handle meaningful workstreams (research, finance, and presentations) using long memory, agent teams, and stronger self-checking. That combination is already rattling cloud/software investors and could reshape how people build software and get office work done.

This article was drafted with the assistance of generative AI. All facts and details were reviewed and confirmed by an editor prior to publication.