Codex on Bedrock can silently bill 2.5–10x: caching never fires
Bedrock users pay full price for a prefix cache that never gets read back; one config flag restores it — plus Codex 0.149 adds an agents dashboard.
Copy markdown
Cache writes every turn, reads near zero
Routing Codex to Amazon Bedrock, it can't mark where its stable prompt prefix ends and rotates its cache key per session — so Bedrock rewrites the whole prefix every turn (cache writes bill at 1.25x input) and rarely reads it back. One analysis pegged the wasted writes at ~2.5x the corrected cost; an affected engineer reportedly saw them reach ~90% of GPT-5.6 spend.
The one-line fix: disable web search
The developer who filed it found setting web_search = "disabled" dropped their bill roughly 10x, because injected search results kept breaking the reusable prefix on every call. Routing through a proxy that inserts the cache breakpoint and a stable key cut costs ~59% in the same writeup, and an open Codex issue tracks a native fix — still unshipped as of 0.149.
Elsewhere: Codex 0.149 adds an agents dashboard
The Aug 20 Codex 0.149 release adds an interactive agents dashboard for juggling tasks, working-directory commands (/cd, /pwd, /cwd), queued messages you can stack on a running session, and expanded codex doctor diagnostics.