Summary: Anthropic shipped Claude Opus 4.6 on February 5, 2026 with agent teams, a 1M-token context window, and Claude in PowerPoint. The article reads this as a move from a lone assistant toward coordinated agents and whole-codebase or enterprise-document workflows. It warns that wider context and multi-agent work also raise cost, observability, and review complexity.
Anthropic shipped Claude Opus 4.6 on February 5, 2026, packing three updates into a single release: agent teams that let several agents work together, a context window spanning 1M tokens, and Claude in PowerPoint.
The date is worth flagging as well — it coincided with OpenAI’s GPT-5.3-Codex. In the opening week of February, the two frontier labs laid their cards on the table face to face.
Agent Teams Retire the Lone Assistant
The structural change is easy to state: a task is no longer shouldered by one model from start to finish. Instead, several agents split the work, coordinate, and fold the results back together — a team you can staff, rather than a single unusually capable helper.
Long-running jobs gain the most: research, implementation, and verification can move at the same time and keep one another honest. The engineering side hardens just as visibly — mistakes travel between agents, costs are harder to rein in, and accepting the final result takes more scrutiny, each tougher to manage with many agents than with one. As capability climbs, operational complexity climbs with it.
For buyers, the live question is what a team of agents is actually for. A single tireless assistant is useful; a team that researches, implements, and reviews in parallel reshapes both the working day and the invoice that follows. Pricing and observability for multi-agent runs are on their way to becoming procurement matters, not engineering footnotes.
What a 1M-Token Context Window Gets You
Once material stretches from lengthy documents to whole-codebase scale, 1M tokens turn the old problem of “look it up first, then condense” into one of “simply place it in context.” Legal work, financial analysis, and reviews of large legacy systems will all have their workflows redrawn as a result.
Context comes at a price: the wider the window, the bigger the token bill and the longer the wait. The question engineers must answer shifts from “will it fit” to “what deserves a place in it” — and filtering plus structuring become more valuable, not less.
A 1M window changes evaluation, too. Material that fits inside a single call eliminates a whole category of chunking and stitching bugs, yet failure gets harder to notice: a model that silently skips page 400 is far tougher to spot than one that never received it. Teams moving large corpora into the window will still want retrieval-style checks, even after retrieval is no longer a necessity for fit.
Agents Set Foot in PowerPoint
Claude in PowerPoint is worth separate attention. Look back at how January set the stage: Claude Cowork (January 12, a GUI agent research preview built for non-coders) and Cowork Plugins (January 30, eleven open-source bundles covering skills, connectors, and sub-agents). PowerPoint support follows the very same path — Claude’s battleground is spreading beyond the chat box into the files and software knowledge workers genuinely rely on.
The edge of what a model can do is moving from “fielding questions” to “acting on files and producing content.” For companies whose operations run on the Office ecosystem, this lands closer to home than any benchmark figure.
It also points Anthropic straight at the document-tooling category. An assistant that can reason about a deck and edit it as well is a different product from a chat window grafted onto a file viewer — and every knowledge-work vendor now has to respond to that direction.