Your agent bill is a context problem, not a budget problem
Every new primitive eventually becomes a bill. Cloud taught us that with compute, storage, egress and GPU hours. Tokens are next, and the first team to hit the wall in public was Uber: their CTO reportedly said the company had exhausted its AI budget months into 2026, largely from coding-agent usage. The instinct is to treat this as a budgeting problem. Set limits, build dashboards, make teams…
Analyze the expenses of AI usage in a business context. The initial thought may be to address it as a budgeting issue, setting limits and monitoring usage. However, the true cause lies in the arithmetic of how tokens are consumed. Anthropic's published rates per million tokens for various models are provided, with Claude Opus costing $5 for input and $25 for output, while Claude Haiku costs $1 for input and $5 for output.
A fleet of five agents, each making 200 calls per day with 20,000 input tokens and 1,500 output tokens per call, results in a monthly cost of $4,125, with 73% of the expense attributed to input tokens. The focus should not be on reducing the volume of tokens written by the model, but rather on understanding that most costs stem from input tokens.
There are four key levers to consider: prompt caching, model routing, batch processing, and reducing context. Prompt caching reduces input token costs by 90% for stable prefixes, model routing can significantly lower costs by switching to a smaller model like Haiku, batch processing offers half-price discounts for asynchronous workloads, and reducing context is the most impactful lever as it lowers the amount of tokens needed, leading to both reduced costs and improved accuracy.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
