What each release does, how to turn it on, and what it showed on my own machine. For every detail, optimaizr changelog --all prints the full list offline.
Claude Code keeps your conversation in a cache for up to an hour. While it is warm, each message re-reads it at the cheapest rate there is. Once it expires, the next message writes all of it again. In the second conversation below that is $1.02 for one message, where re-reading it costs $0.025. 0.10 shows you which one you’re in, before you send.
The conversation is cached. Each message re-reads it cheaply, and the minutes are how long that lasts if you stop now.
It counts down, and says what coming back after will cost. Leaving? Ask for a handoff note, then /clear.
Your next message writes the whole conversation again. Starting a new task? /clear first, and it starts small.
Your 5-hour window, at the end of the line.
The chat panel draws no status line. Run optimaizr live in a terminal beside it for the same countdown.
Every message re-reads the whole conversation, so the longer it runs, the more of each message is re-reading. sessions measures where that happens in your own logs, with nothing estimated.
Counts down too, warns 5 minutes before an expensive cache expires.
A new rule, for Claude Code and Codex, priced against a fresh start.
/optimaizr handoff: a short note for cents, and after /clear the next conversation starts from it.
Its band counts down the last 15 minutes of the cache.
Compacting prices the reload at the 1-hour rate your conversations use.
Grouped: start here, change things, go deeper.
Before 0.9 a finding was one number. Now each saving says how sure it is: clear waste you can drop with nothing to lose, likely savings if you try them, and levers that save up to an amount, never added together. On my machine that day: $13.07 a month of clear waste, and up to 28% less by compacting long conversations earlier.
Your Claude plan is read from Claude Code itself. No --plan.
Clear waste, likely savings and levers, each labelled, never added up.
Every call re-reads the conversation. Compact earlier, undo any time.
A status line, context warnings, and Y to compact.
Its own settings and compactions, GPT-5.4 to 5.4 mini.
Whole tasks, not single calls. Numbers that hold up.
Claude Code runs mods. Press Y in optimaizr live, or let --auto do it, and the session you’re in moves to a cheaper model when that pays. The HUD shows what it saved. The mod is optional: everything else works from the terminal.
Y, or --auto, switches the session you’re in, when it pays.
Turn cost and your 5h window, live. /optimaizr hud for the rest.
What it cost, and what a switch saved, in dollars and of your window.
Lower effort, same model.
A command that failed twice, held once.
/optimaizr off goes back.
--plan team and --plan team-premium show 5-hour sessions for Claude Team seats; every ChatGPT plan Codex reports is named and priced.verify's judge no longer runs out of room before its verdict.Y in live and apps using wrap() switch model from their very next request, no restart; optimaizr undo <rule> takes it back.--budget 300 names the day a monthly cap runs out, and how many days the fixes buy back; live warns at 50, 80, 95 and 100%.--plan pro for Claude, with waste per 5-hour window.optimaizr card turns your last 30 days into an image to post: totals only, no project names.optimaizr profile puts spend, waste and your single biggest opportunity on one screen, with the next command to run.