GAIA snapshot shows scaffolding, not the base model, drives agent benchmark scores

Price Per Token·medium signal

A July 2026 GAIA snapshot underlines that agent benchmark scores depend far more on scaffolding than on the base model: Princeton's HAL 'scaffolded' leaderboard shows Claude Sonnet 4.5 at 74.6%, while the 'bare model' board tops out at GPT-5 Mini's 44.8%. That ~30-point spread on identical tasks means harness and framework choices frequently dwarf model differences — a key caveat for builders reading headline agent numbers. Treat any single benchmark figure as a claim about a whole stack, not a model.