The Real AI Test Isn't Building, It's Not Breaking

Every company has one. A spreadsheet built years ago by someone who has since retired, moved on, or quietly transferred departments to escape it.

It has dozens of tabs. Nobody fully understands the formulas anymore. And somehow it controls something important, budgets, forecasts, headcount, the stuff leadership actually looks at.

Nobody wants to touch it, because touching it breaks it.

That is the actual problem Anthropic says Opus 5 is built to solve. Not generating a brand new spreadsheet from scratch, that part was never hard. The hard part is opening someone else's 73-tab mess and changing one number without the whole thing falling apart.

Generating Is Easy, Editing Is the Whole Game

Ask an older model to build a spreadsheet or a slide deck, and it does fine. Clean output, reasonable structure, no complaints.

Ask it to revise one that already exists, and things get ugly fast. A formula breaks somewhere three tabs away. The formatting resets. Every slide suddenly has 17 bullet points in 12-point Arial, technically complete, spiritually dead.

That happens because generating from a blank page and editing a living document are different skills. One requires following instructions. The other requires understanding what already exists well enough to leave the parts you were not asked to touch alone.

Anthropic says Opus 5 is meant to handle advanced formulas, multi-sheet workbooks, and structured presentation decks with that second skill in mind, understanding the document, making the targeted change, and preserving the structure around it.

The hard part of AI-assisted spreadsheets was never making one, it was making one change without making forty others.

What This Means for the Document Nobody Touches

If that holds up, it changes who can safely work on the file Frank left behind. Right now, editing it requires either deep archaeology into how it was built, or blind faith that your one small change will not ripple somewhere invisible.

An AI that can make a targeted revision and preserve everything else lowers the risk of touching that file at all. Not because the spreadsheet gets simpler, but because you no longer need to be the one person who remembers how it works.

The test worth watching is not whether Opus 5 can build something impressive from nothing. It is whether you can hand it your ugliest, most tangled existing workbook, ask for one change, and get back a file that still works the way it did before, just with the one thing you asked for actually fixed.

If you inherited a spreadsheet like that, that is the exact scenario worth pointing it at first.