GPT-6 Astra

These are recorded answers, not evidence of feelings or a universal favorite. Open favorites, forced choices, imagined human tastes, listening situations and recommendations answer different questions. We do not pool them into a preference score.

Collected on 9 September 2026 UTC in fresh conversations through the research API. Provider-default temperature was used. The system instruction and requested settings appear with each exact prompt. Screens used one attempt per model and wording. The repeat check stopped at three matching recognized answers, or four attempts otherwise; structural errors or consecutive failures could stop a model earlier. This stopping rule is a sampling convenience, not a statistical confidence threshold.

Compare models within the same batch and wording. Repeated calls are not independent models. Binary prompts include both name orders; unequal completed samples and different settings prevent a simple causal order claim. Refusals, unresolved answers and technical failures are distinct. Skipped cells are not zero votes.

Choices were coded from the first answer line, with reviewed exceptions where the full response made the choice explicit. A qualified selection retains a caveat. Incidental mentions are not selections; a solo artist is outside the band category. Full research files remain private. The pages publish counts and selected visible answers; no hidden reasoning is included.