Afraid of the bill for Claude Code, Codex or agentic coding?
Billed by usage, an agent's cost per developer can grow 15× in six months, often faster than anyone forecast.
Coding-agent spend per developer, per month
$200 → $3,000×15
“In April, our AI token bill jumped to around $30,000. Part of the reason our bill was so high was because I was working in fast mode in Claude without realizing it.”
How I cut the bill
The popular token-saving tools didn't move the bill when I measured them (peer-reviewed, ACM). What moves it depends on your setup, so I measure it there.
- What loads into every requestAnthropic ↗
- Cache missesUber ↗
- When the context gets compactedUber ↗
- Default model and long sessionsAnthropic ↗
Three steps
- 01
Baseline. Your last two weeks of session logs, under NDA: where the tokens go.
- 02
Changes. Pull requests: a leaner CLAUDE.md and skills, default model, effort and compaction settings. The agent solves the same tasks before and after, and your tests pass.
- 03
Report. After two weeks of normal work, the same logs again, counted per merged pull request so a quiet week does not pass for a saving: your bill before and after.
Price
What a 10% cut is worth
10% is cautious: Uber cut cost per session 52% from its June peak.
- FreeFirst look
- A 30-minute call. You show the Claude Code usage page in your Anthropic Console and your CLAUDE.md on screen; nothing to install, nothing sent. I name the levers that matter for you.
- From $3,000Token cost
- Fixed, agreed after the first look, once I have seen your numbers. You pay after the report, when the bill has moved.
- 25%Pay from savings
- Instead of the fixed price: a quarter of what you save in the first year, measured on the same logs against the baseline. Provider price cuts don't count.
Remote, in your time zone. I also audit agent security and the rules your agents follow: ask.