We cut Codex's input token cost by 49.5% with a compression gateway (benchmark) (edgee.ai)

No discussion on problem difficulty, or on result quality besides "the Edgee run generated slightly more output tokens than the baseline".

what part do you compress more specifically ?