Source note
Token spend breaks budgets – what next?
Summary
The article reports that AI coding-agent token costs have risen fast enough to force budget decisions at many software companies. It matters because agent use is moving from a developer tool cost to a visible engineering operating expense.
Problem
- Companies want engineers to use coding agents, but token spend can grow faster than budgets and planning cycles.
- Expensive model defaults, high-effort settings, and uncapped agent loops can turn normal coding work into hundreds or thousands of dollars per developer.
- The main management problem is deciding when higher spend pays for faster delivery and when it is waste.
Approach
- The author interviewed developers and engineering leaders at 15 companies, with examples from large firms, mid-sized firms, startups, and seed-stage companies.
- The article compares spend patterns, tool choices, default models, caps, discounts, and management responses.
- The core mechanism is simple cost control: measure token use, route simple tasks to cheaper models, set cheaper defaults, cap spend, or accept higher spend when output rises.
- The reporting separates two common responses: keep usage open while measuring impact, or curb spend through model choice and budget limits.
Results
- Two companies reported token spend rising about 10x in six months, with no clear slowdown.
- A seed-stage AI infrastructure company reported a 15x rise in six months, from about $200 per developer per month to about $3,000 per developer per month across seven developers.
- One fintech reported some developers spending $500 per day on Claude Code, with employee costs effectively doubling and code review becoming the bottleneck.
- A 2,000-person SaaS company said changing the default model cut costs by 30%.
- A finance company with about 800-1,200 users each on Cursor and Claude Desktop found that a $100 per-user limit could be exhausted in 3-5 working days.
- Other concrete figures include a $10K one-week caching mistake, high-end users around $1K per week, a $1,400 single-day Claude Code session, traffic growth above 10x year over year with the same engineering team, Cursor discounts after about $1M in spend, and no Anthropic discounts reported by companies spending $5M+ per year.