Shift-1 vs. top-model baselines

Shift-1 holds quality close to an all-top-model setup at 45% and 76% of the cost

HARNESS Codex Claude Code
Shift-1 vs. top-model baselines
There are two reasons Shift-1 is better on Codex than Claude Code. Shift-1 is trained/tuned primarily on Codex conversations. Claude Code has more modest gains on cost due to the extremely short cache timeout (5 min) and lower cost differential between Sonnet/Opus.