Fable 5.1 in Low mode: the two-step workflow for Excel, PowerPoint and Word

By Nate (Nate's Substack) · September 4, 2026.

Nate explains that Fable 5.1 has changed his mind about how a frontier model should be used for knowledge work (Excel, PowerPoint, Word). This email is the written version, a companion to his video review of the same model, focused on how to apply it in practice rather than on what it is capable of doing.

As a test, he asked Fable 5.1 to research a pending merger between GoPro and Starman Optical, build a post-close discounted cash flow (DCF) model in Excel, and turn that analysis into a PowerPoint presentation. Running it on the cheapest setting ('Low'), from a single short prompt, the model produced a seven-tab Excel workbook with working formulas and scenarios that actually moved the valuation, plus a 13-slide deck that, according to Nate, was already usable.

His recommended workflow is to always start on Low so the model builds the spreadsheet, the deck or the draft, and only then open the output and move up to the 'High' or 'Extra' settings for the specific part that requires human judgment. The reason, he explains, is that a Low run is already good enough to reveal what is wrong, so you spend the expensive tokens after you understand the problem, not before.

Reviewing the generated files, Nate noticed that a 'Sources' tab and a 'Checks' tab were missing: he could follow the analysis, but he struggled to trace where the model's assumptions came from. That, he says, was the point where a run on the 'Extra' setting made sense, and where another model he calls 'Sol' turned out to be better at that specific part of the job (no further details about Sol are given in the body of the email).

Nate argues that this 'start cheap, level up only where needed' pattern matters more in knowledge work than in programming, because code offers clear feedback (the program runs, the test passes, or an error points to what failed), whereas a Word document, an Excel sheet or a presentation gives no warning when an argument is weak, and a spreadsheet will calculate a wrong assumption with total precision all the same. For him, the only way to gain clarity is to work directly on the file: building the sheet changes what he understands about the business, building the deck changes the story he thinks needs to be told, and writing forces him to decide what he actually thinks. His conclusion is that you discover the story as you go, and that Fable 5.1 fits that process well because its Low setting is competent enough to make each iteration useful and cheap enough to keep iterating without paying the price of the most expensive setting on every turn.

The email previews the table of contents of the full article, which promises to cover: exactly what Fable 5.1 produced on Low (including the two tabs it left out and where Sol excelled); what the model's five effort levels really change, what they cost, and a warning about using the Low setting for research tasks; the specific prompts Nate would use for Excel, PowerPoint and Word, including a particular instruction to stop Fable from rebuilding an entire workbook when only a small change is requested; his conclusions on Fable 5.1 as a writing tool, based on a 100-word test comparing three models, and how to tell useful edits from those that flatten your own voice; the instructions he adds before launching long jobs that have to finish unsupervised, citing as examples a 37-second architectural short film and a short about 'Microduck'; and finally a guide gathering all the prompts in one place.

The body of the email cuts off just before developing these points ('Let's start with what Low actually produced...'), indicating that the full treatment, along with the prompt guide, is reserved for paying subscribers of Nate's Substack.