Anthropic drops Claude Opus 4.6 with 1M context

While most people are still using AI as an over-cautious drafting assistant, Anthropic wants its models running entire analyst desks.

On Thursday, Anthropic rolled out Claude Opus 4.6, refreshing its flagship model with deep agentic coding upgrades, a massive 1M-token context window in beta, and a direct push into Microsoft Office workflows.

Multi-agent squads and fixing context rot

The marquee engineering update is agent teams in Claude Code. Instead of a single model plodding through a monolithic codebase sequentially, you can now spin up parallel subagents that split tasks, coordinate work, and read repositories concurrently. Developers can jump in and steer any subagent directly using Shift+Up/Down or tmux.

Anthropic also tackled one of the most frustrating aspects of heavy AI usage: context rot. When models swallow hundreds of thousands of tokens, their attention usually falls off a cliff. On the rigorous 8-needle 1M MRCR v2 retrieval test, Sonnet 4.5 managed just 18.5%. Opus 4.6 jumped to 76%.

To keep long workflows from choking, Anthropic added context compaction on the API, which quietly summarizes historical turns as conversations hit limits. Developers also get adaptive thinking and four effort tiers, ranging from low to max, meaning you no longer have to pay high-latency reasoning costs on trivial logic.

Moving from terminal windows to boardrooms

Anthropic noticed that non-developers were flocking to Claude Code simply because its autonomous execution loop was effective. With Opus 4.6, the company is meeting those knowledge workers where they already live.

On GDPval-AA, which measures economically valuable knowledge work across finance and law, Opus 4.6 cleared OpenAI's GPT-5.2 by roughly 144 Elo points. It also posted 60.7% on Vals AI's SEC-filing benchmark Finance Agent and 76.0% on TaxEval.

Those numbers power practical desktop tools. Claude in Excel now handles financial-grade formatting, pivot table edits, charts, and conditional formatting without user hand-holding. Claude in PowerPoint arrived as a side-panel research preview, letting users refine slide masters and decks natively rather than exporting raw decks from a chat window.

For enterprise desktops, paid users can test Cowork, an environment that lets Claude read, edit, and create files inside designated folders alongside custom plugins for reconciliations and variance analyses.

Base pricing stays at $5 per million input tokens and $25 per million output tokens, though prompts stretching beyond 200k tokens jump to a premium $10 and $37.50 tier.

Why it matters

Anthropic is making an aggressive and very calculated pivot from pure code generation to white-collar operational work.

For investment bankers, corporate finance teams, and senior developers, Opus 4.6 is immediately worth testing on stubborn, multi-step workflows. Commercial due diligence tasks that typically tie up junior analysts for weeks can now receive legitimate, structured first-pass models directly in Excel. The trade-off is cost: running multi-agent swarms over a million tokens hits the higher billing threshold rapidly, meaning careless experimentation will burn through API budgets fast.

Across the industry, this escalates the enterprise proxy war against OpenAI. Just as OpenAI rolled out its Frontier agent platform and GPT-5.3-Codex, Anthropic bypassed abstract agent infrastructure and embedded Claude directly inside the apps where corporate budgets actually live. Turning Claude into a native Office co-worker is a brilliant distribution move, but letting autonomous subagents modify spreadsheets and local directories still carries compounding error risks if human verification slips.