The true cost of saying "Hi" to an AI agent

7 hours ago

AI agents often overthink simple greetings like 'Hi', leading to high time costs despite low token costs.
Benchmarking 14 models shows that waiting time, not token pricing, is the real expense; waiting can cost over 20 times more than tokens.
Some models (e.g., Sonnet) respond to 'Hi' with excessive tool calls, auditing repositories and making unsolicited commits.
Failure rates occur with ambiguous prompts; models like Haiku and MiniMax failed in multiple runs when greeted.
Clear tasks (e.g., 'commit') are handled efficiently by all models, highlighting that ambiguity drives up costs.
The total cost of interaction combines API fees and waiting time based on salary, making latency a key factor.
Optimization advice: Be direct in prompts to avoid unnecessary delays, as ambiguity increases costs significantly.
The benchmark used OpenRouter with 210 trials, emphasizing the need for faster, cheaper models to reduce overall costs.

Hasty Briefsbeta