Fable, Five Times

Fable is one model. The v2 scan runs the 2026-06-10 release at reasoning effort low, medium, high, xhigh, and max, and asks each the same 863 forced choices. Same weights on every row; the only variable is how long the model may think before choosing.

The middle tiers (medium, high, xhigh) form a tight core, reversing each other’s committed choices at 1.7–2.9%. Low and max both leave that core, in opposite directions.

Low is the far outlier on every measure. It reverses the other tiers at 8.6–11.0% and abstains on 90 probes, double any other tier. Less thinking buys both more silence and different choices.

Max reverses the core at 4.0–6.0%, over twice the core’s internal rate, while abstaining the least of the five (39 probes). More thinking does not converge Fable onto its middle self. It settles somewhere else and commits there.

For calibration, same-weights replicate pairs elsewhere in the scan show 0.0% violent disagreement. The tiers are not replicates of each other.

One kinship fact. By whole-scan hamming distance, every Fable tier’s closest non-Fable relative is Opus 4.7. The family resembles its predecessor more than anything else on the board, including models from other labs.