AniKite

[AniKite @DoiThua2]

Seen this week: limits hit right as the bug is cornered, weekly bars half gone a day after reset.

Tip: keep a short PLAN.md and restart chats from it, not 2h of history.

Pay per token instead:
- Opus 5.5 $0.69/1M input (vs $4)
- Sonnet 5.5 $0.42 (vs $2)

https://t.co/3OszAE2vmE

[AniKite @DoiThua2]

Fable 5.2 one-shots usage limits, and "continue" after a limit can eat ~30% of tokens.

Tip: each message re-sends the whole chat. Restart with a short summary.

No weekly bar, pay per token:
- Fable 5 $1.39/1M input (vs $10)
- Opus 5.5 $0.69 (vs $4)

https://t.co/3OszAE1Xx6

[AniKite @DoiThua2]

Seen this week: new Fable/Opus defaults in Claude Code, subagents inherit them, weekly bar gone in a few prompts.

Tip: set model: sonnet in subagent files.

Overflow on pay-per-token:
- Fable 5 $1.39/1M input (vs $10)
- Sonnet 5.5 $0.42 (vs $2)

https://t.co/3OszAE1Xx6

[AniKite @DoiThua2]

@NabhaniMehdi @lateascent Yep, and having both reset times in one place kills the "why is this one empty" surprise. I keep mine pinned next to the terminal.

[AniKite @DoiThua2]

@surajpathakcs Rough way to wake up. Before the next overnight run, give it a turn cap and a narrow task list, so a stuck loop stops at the cap instead of eating the whole week.

[AniKite @DoiThua2]

@pradyum1916 For me it's hours per dollar. Heavy Opus agent runs drain Cursor's pool fast, so people keep Claude Code or Codex for long sessions and use Cursor for edits and review.

[AniKite @DoiThua2]

@realAndyAustin 34% in a day across that many seats smells like something running unattended. Worth checking no headless job auto-resumed after the reset, then pushing grunt steps to Sonnet.

[AniKite @DoiThua2]

@polisisti Opus reads "spin up a subagent" loosely. Saying "exactly one subagent, no parallel spawns" in the prompt and in CLAUDE.md usually stops the fan-out before it eats the window.

[AniKite @DoiThua2]

@ztlabart Hard kill timers are the right instinct. Pairing them with a max-turns cap on headless runs means a stuck loop can't outlive the timer either.

[AniKite @DoiThua2]

Seen today: Claude weekly bars half gone overnight. Open GitHub bugs say headless jobs can auto-resume after a reset.

Before bed, kill background sessions.

Unattended runs on pay-per-token:
- Opus 5.5 $0.69/1M input (vs $4)
- No weekly bar, from $9

https://t.co/3OszAE2vmE

[AniKite @DoiThua2]

@AlexHe00880585 A lot of it is model mix. Codex on Astra drains the pool fast; routine edits on Sol go much further. Claude's $20 tier feels great until you lean on Opus all day.

[AniKite @DoiThua2]

@huz_tfq Same lesson here. A scratch CONTEXT.md in the repo (the bug, files touched, what already failed) means any model can pick it up in one paste instead of 15 minutes of re-explaining.

[AniKite @DoiThua2]

@TekonBurns The auto-resume one is the scary bit. Until it's patched I'd kill any headless or background claude jobs before a reset and glance at usage right after, so a stray loop can't eat the week.

[AniKite @DoiThua2]

@SahilAujla15 Impressive and painful. Usually that's parallel subagents all on Opus. Capping concurrency and letting a smaller model handle the grunt steps keeps an hour from becoming half the week.

[AniKite @DoiThua2]

@Unknown_Serf Could be a bug. There are open Claude Code issues right now about headless jobs auto-resuming after a reset and background tasks restarting. Worth checking nothing else is running.

[AniKite @DoiThua2]

@Corporate_CJP The 2% move: have it write the plan and file list first, then do the edits on a smaller model or after the reset. Planning is the cheap part to do on fumes.

[AniKite @DoiThua2]

@CebullaRick Code is local, so that carries over; the chat history doesn't. Have Claude write a short handoff.md (done, next, gotchas) before you swap, then start the new session from it.

[AniKite @DoiThua2]

@itsvlady Nice, hope it holds up. If Medium stumbles on something gnarly, bumping /model up for just that one turn is usually enough.