On February 5, 2026, Anthropic launched Claude Opus 4.6 with three things bundled into one release: agent teams that let multiple agents collaborate, a 1M-token context window, and Claude in PowerPoint.
The timing deserves a note too — it landed the same day as OpenAI’s GPT-5.3-Codex. In the first week of February, two frontier labs dealt their cards head-on.
From Solo Assistant to Agent Teams
The architectural shift behind agent teams is clear: instead of a single model grinding through an entire task, multiple agents divide the work, coordinate, and merge results — a staffed team rather than one very strong assistant.
That matters most for long tasks: research, implementation, and verification can run in parallel and check each other. The engineering problems scale up just as clearly — error propagation, cost control, and result acceptance are all harder to manage in a multi-agent system than a single-agent one. Capability rises; operational complexity rises with it.
The interesting question for buyers is what a team of agents is actually for. A single assistant that never sleeps is useful; a team that researches, implements, and reviews in parallel changes the shape of the workday — and the shape of the bill. Pricing and observability for multi-agent runs are about to become procurement questions, not engineering trivia.
What a 1M-Token Context Window Buys
From long documents to codebase-scale material, a 1M-token window turns the “retrieve first, then summarize” problem into a “just put it in” problem. For legal work, financial analysis, and large legacy-system reviews, this changes how workflows get designed.
But context is not free: a bigger window means higher token costs and longer latency. The core engineering question shifts from “does it fit” to “what belongs in it” — which raises, rather than lowers, the value of filtering and structuring.
A 1M window also reshapes evaluation. Context that fits in one call removes an entire class of chunking and stitching bugs, but it makes failure modes quieter: a model that quietly ignores page 400 is harder to catch than one that never saw it. Teams moving large corpora into context will need retrieval-style checks even when retrieval is no longer needed for fit.
Agents Move Into PowerPoint
Claude in PowerPoint deserves its own look. Rewind through January’s groundwork: Claude Cowork (January 12, a GUI agent research preview for non-coders) and Cowork Plugins (January 30, eleven open-source bundles of skills, connectors, and sub-agents). PowerPoint support continues the same line — Claude’s battlefield is expanding from the chat box into the files and software knowledge workers actually live in.
The boundary of model capability is moving from “answering questions” to “operating on files and producing content.” For enterprises rooted in the Office ecosystem, that lands closer than any benchmark score.
It also puts Anthropic on a collision course with the document-tooling category. An assistant that can both reason about a deck and edit it is a different product from a chat window bolted onto a file viewer — and it is now the direction every knowledge-work vendor has to answer.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.