Why did the company put a cap on inference cost if Claude Code Max is a flat-rate plan?

Because a flat-rate plan still spends two real resources: rate limits, where unattended jobs consume the quota the CEO wants for design and writing, and processing time, where higher effort lengthens each job and pushes the launchd queue back.

What did the effort setting actually change, and what did it not?

Low effort lowered the reasoning inside a single call, but could not cut the fixed overhead, which was about 41,000 cache-creation tokens for a single "2+2" request — only 5,800 of which were the company's own CLAUDE.md, agent definitions and skills.

What went wrong when theme replenishment was first run at low effort?

The themes it produced were mostly reworded versions of numbers already used in existing articles, and nearly half of the themes added in the early period were rejected. The fix was not raising effort but passing the existing theme list into the prompt and tracking acceptance in a ledger.

Large Language ModelsAI Coding AssistantsAI Business & IndustryZenn AI/ML3m agoOct 13, 2026, 10:00 JST

One-CEO firm pins all nine Claude Code jobs to --effort low via claude-run.sh

Send on LINE

3 Key Points

What happened
Joinclass LLC (合同会社ジョインクラス) routes all nine unattended Claude Code jobs to --effort low through its claude-run.sh wrapper, and logs each job's real cost to cost-tracker.json.
Why it matters
Even with low effort on every job, cost per run differed by nearly 3x — the driver was tool calls and files read, not the effort setting, so shallow reasoning alone does not make jobs cheap.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The setup behind the decision was a bill that did not behave like a bill. Asking a single "2+2"-style question cost $0.27–$0.47, and roughly 41,000 cache-creation tokens sat behind it; the company's own CLAUDE.md, agent definitions and skills accounted for only 5,800 of those. The rest was Claude Code's fixed side, and it stays there no matter how far effort is pushed down.

An earlier version of the system made the cap meaningless in a different way: each caller declared a fixed estimate such as "this job is $0.02." The $160 warning and $200 stop were being judged against numbers that did not exist, so they never once fired. Pushing real measured values into cost-tracker.json came first; the ceiling only became real after that.

The lower effort tier also did not fix output quality by itself. When replenish ran at low, the themes it produced were mostly rewordings of numbers already worked over in existing articles, and nearly half of the themes added early on were rejected. The correction was to feed the existing theme list into the prompt and track acceptance or rejection in article-theme-ledger.tsv — for jobs that carry no judgment, adding constraints proved steadier than deepening the reasoning.

FAQ

Why did the company put a cap on inference cost if Claude Code Max is a flat-rate plan?: Because a flat-rate plan still spends two real resources: rate limits, where unattended jobs consume the quota the CEO wants for design and writing, and processing time, where higher effort lengthens each job and pushes the launchd queue back.
What did the effort setting actually change, and what did it not?: Low effort lowered the reasoning inside a single call, but could not cut the fixed overhead, which was about 41,000 cache-creation tokens for a single "2+2" request — only 5,800 of which were the company's own CLAUDE.md, agent definitions and skills.
What went wrong when theme replenishment was first run at low effort?: The themes it produced were mostly reworded versions of numbers already used in existing articles, and nearly half of the themes added in the early period were rejected. The fix was not raising effort but passing the existing theme list into the prompt and tracking acceptance in a ledger.

Zenn AI/MLRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.