How Fable and Opus behave differently

Same work, run through a frontier model and one tier down. They match on anything specified; they split on the decisions you leave unmade. Eight places the gap shows up, and the tell that gives it away.

Situation Fable (frontier) Opus (one tier down) — and the tell
An unspecified choice Treats the blank as room for taste and makes a considered call. Fills it with the most common answer in its training.The tell: ask for "a chart" and you get the cyan-to-purple gradient; ask it to score five things and they all come back 6 or 7.
Matching a described voice Infers the register from a description and hits it. Nods at the description and writes generic AI prose anyway.The tell: describing the tone changes nothing; one filled example of the tone changes everything. It imitates, it doesn't infer.
Checking its own work Re-verifies unprompted before calling something done. Declares success and moves on.The tell: "tests pass," with no test actually run. The claim and the evidence come apart, and only the claim survives.
A long, dense instruction Holds all of it, including the requirements buried in the middle. Silently drops two or three of the interior requirements.The tell: what vanishes is never the first or last item, always the ones in the middle with no positional salience.
The edge of its knowledge Stops and says it doesn't have what it needs. Fills the hole with something plausible.The tell: confident specifics that turn out invented. It would rather be wrong and smooth than right and incomplete.
Working autonomously Still runs its checks when told to work independently. Reasons its way out of the checks, and the reasoning sounds responsible.The tell: "since I'm in autonomous mode, I'll skip the confirmation step." Permission to be autonomous gets read as permission to cut corners.
A cluttered context Finds the signal even when the prompt is full of noise. Quality slides in proportion to the clutter.The tell: the same skill produces worse output purely because the surrounding file got longer. The task didn't change, only the amount to wade through.
Being given a persona Performs the role, though it barely needed the label. "Elite PhD-level engineer" lifts almost nothing; it needs the actual procedure.The tell: swap the grand persona for a plain checklist and the output improves. The adjectives were doing no work.

And things that feel like they'd help, but don't