Testing GPT-6 Astra for image labeling: magical until you see the usage bill

chris_j_paxton · x · 2026-09-09

Matt Franco tried GPT-6 Astra on a batch of images after the viral basketball labeling demo — it felt like magic until he saw the usage. He concludes battle-tested HF models are still the safer bet, but found a hybrid workflow promising: let Astra manage the runs and step in manually to correct labels.

Original post →

More from coding & agent

coding & agent channel →