Caching was the theme of the day, and the numbers were less exciting than the announcements. Elsewhere, a security scan that cost real money and found nothing, a review queue that needed a cap, and a Go team that stopped letting its agent grep.
01
// usage bill, failure
A $250 Codex security scan that returned 0 findings
$250 charged, two weekly limits used, 0 findings
A builder ran Codex security scans and used up their weekly limit, which led to a $250 charge. They then used a reset, and the scan burned through the weekly limit a second time. When they asked it to show the results so far instead of continuing, to save at least some tokens, it cancelled the run and returned 0 findings. Their verdict on the scan UX: "horrible".
After OpenAI shipped a 12-hour prompt cache for long-running agents, one team replayed a week of their own Claude Code logs (73,281 calls in 2,320 sessions, Oct 14 to 20) against it. 98.5% of input tokens already came from Claude's 1-hour cache, and only 38 calls (0.05%) came back after an idle hour, so a 12-hour cache would have cut the input bill by 0.4%. The real cost was elsewhere: 61% of their cache writes landed within 5 minutes of the previous call, and 34% were the first call of a session. The author builds an agent harness product (Tectomight), so this is their workload, not yours.
95.4% cache hit rate, $537.72 still spent on rebuilds
$537.72 in 30 days on cache rebuilds
A research engineer checked 30 days of their Claude Code usage and found a 95.4% cache hit rate, which looks healthy. Yet $537.72 of input spend, about a fifth of it, still went to rebuilding the cache. Their point: hit rate counts tokens, while the bill counts misses times context size, so a few misses on a big context cost more than the percentage suggests.
One developer ran three Codex worktrees in parallel. They opened 9 PRs before lunch, and the developer had reviewed 4 by dinner. The agents were not the bottleneck, the human review was, so they added a cap.
Give each agent a limit of 2 open PRs, and make the third task wait until you merge one.
05
// fix shipped
The agent grepped for callers and got 146 lines for 3 real ones
146 lines of search output, 3 real callers
A team asked a coding agent who calls a function in their Go backend. It searched the text and came back with 146 lines, for 3 real callers. Their fix was to let the agent ask gopls, the Go language server, which answers from the compiler instead of from string matches.