GPT-6 against Fable 5.1 for first drafts: what actually changes

A new model arrives, the benchmark charts go round, and none of it tells you whether your Tuesday gets easier. Benchmarks measure narrow tasks under conditions nothing like your inbox. The only comparison worth your time is the one you run on your own work.

Here is how to do that in an afternoon, and what tends to separate the two in practice.

Build a test set from work you have already done

1. Collect five real tasks you did last month. A client email, a product description, a summary of a long thread, an outline, a rewrite.
2. For each, keep what you actually sent as the standard to beat.
3. Write one prompt per task, as you would normally write it, and use the same prompt in both models.
4. Judge only two things: how much editing the draft needs, and whether anything in it is false.

What tends to differ in real drafts

Length discipline. One model will give you the six hundred words you asked for and the other will give you nine hundred and a summary you did not want. Whichever way round it falls for you, it is the single biggest source of editing time.

Instruction stacking. Ask for a tone, a length, a forbidden word and a required call to action at once, and see which constraints survive. Most disappointment with a model is one of four instructions being quietly dropped.

Confident invention. Ask a question about something the model cannot know, such as a private price list or last week's policy change, and see whether it says so. A model that invents a plausible answer costs you more than one that writes slightly duller prose.

The prompt structure that lifts both

• Say who it is for and what they already know.
• Give one example of the voice you want, from your own writing.
• State the length as a number of words and the format as a list of sections.
• Name what must not appear: the phrases, the claims you cannot support, the sales language.
• Ask for the draft only, with no preamble and no closing offer of help.

After the draft, the boring part still matters

Both models produce text that is longer than it needs to be and fond of a few patterns: long dashes, the same three word rhythm, headings that promise more than the paragraph delivers. Cleaning that up is mechanical and fast if you do it deliberately.

Clean up pasted text

What not to conclude from the comparison

That the winner is the winner for everything. Models differ by task: one may be better at summarising a long thread and worse at a sales email. Keep the score per task rather than as a single verdict, and rerun the test when either is updated, which will be sooner than you expect.

Common questions

Does a longer prompt always produce a better draft?

No. Beyond a point extra words compete with each other and constraints get dropped. Precise beats long.

Should I use one model for everything?

Most people settle on one for drafting and another for checking, because reading your own draft with fresh eyes is exactly what a second model is good at.

How often should I rerun the test?

When a major version changes, or when you notice yourself editing more than you used to. The test set is written once and reused.