How much does an AI agent cost to run? Numbers from 4 builds (madsalt.tech)

In a recent release, AI engineer Denys shared insights into the operational costs of running AI agents, revealing the three primary expenses: model calls, infrastructure for system operations, and necessary guardrails to prevent failures. The analysis centered on four AI agents built over the year, highlighting significant disparities in model call costs — ranging from $0.00046 for DeepSeek V4.1 to $26 for Claude Opus 5.5 per thousand steps. Interestingly, most costs derived not from model calls themselves but from operational infrastructure and responses to failures, such as unnoticed outages that can extend for days. This detailed breakdown is significant for the AI/ML community as it emphasizes the importance of optimizing not just the choice of models but also the supporting systems that enhance reliability and cost-efficiency. Key takeaways include the strategy of routing messages to the cheapest competent model, the necessity to limit unnecessary model calls, and the potential for on-device processing to bypass API costs entirely. Denys's findings reinforce that managing context and operational surroundings are just as crucial as the models chosen, providing valuable lessons for developers engaged in AI agent projects.