This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
Anthropic released Claude Sonnet 4.5 yesterday, and most of the launch post is about agents and code. By Anthropic’s measures it is state of the art on SWE-bench Verified, leads the OSWorld computer-use benchmark at 61.4%, and in testing has kept focus “for more than 30 hours on complex, multi-step tasks.”
The line I keep returning to is a quieter one: “In the Claude apps, we’ve brought code execution and file creation (spreadsheets, slides, and documents) directly into the conversation.” The feature was previewed for some plans earlier this month. It is now on all paid plans.
The usual reading is that this is a productivity story, and that a better model simply means better decks. For anyone who cares how a company is described, I think the more important change is not in the model at all. It is in what leaves the chat.
The file travels alone
A chat answer comes with its own context. Someone typed a question, read the reply, maybe noticed a hedge or a source link and asked a follow-up. A file carries none of that. In the September preview, Anthropic described a private computer environment where Claude writes and runs code to produce files, and Engadget noted that users download the result once they are happy with it. From that point, the slide deck about a competitor, the briefing document on a supplier or the market summary that mentions your company is just a file. It gets renamed, attached, dropped in a shared folder and opened by people who never saw the prompt.
The forwarder becomes the source
This is the part I think gets missed. When a colleague sends a deck, the recipient does not judge the chatbot. They judge the colleague. The authority of the document quietly transfers to whoever forwarded it, and that person may have spent a few minutes checking the layout and none checking the facts.
So an outdated description of a company, a lawsuit attached to the wrong business, or an executive listed in a role she left a year ago does not circulate as an AI answer. It circulates as something a trusted person produced. That is a stronger position for an error than any chatbot screen.
Conversion compresses the record
Anthropic’s own examples include turning a PDF report into slides and meeting notes into a formatted document. Conversion is where nuance goes missing. A paragraph explaining that a regulatory inquiry was opened and later closed can become a slide titled “regulatory issues.” Every individual step looks like summarizing. The end product is a new document that says less, with more confidence.
Better models do not close the gap
None of this is a complaint about Sonnet 4.5. Anthropic calls it its most aligned frontier model yet and reports progress against prompt injection. But a stronger model changes how often mistakes are made, not what happens to the ones that remain once they are formatted, branded and sent.
What follows
For communications and investor-relations teams, assume that analysts, buyers, journalists and competitors will increasingly hold AI-built documents about your company that you will never see. The defense is upstream: dated, plain-text facts on your own site about leadership, structure, results and how past incidents ended, so a tool assembling a file has something current to find.
Inside your own organization, treat any generated deck or document that names a real company as a draft whose facts are unverified until someone checks them against a primary source. Asking the same assistant whether its own file is right is not that check.
A chat answer at least looks like something a machine said. A forwarded deck looks like something a colleague did.