I recently set up Hermes Agent on my Mac mini with Claude Sonnet 5 as its main model through OpenRouter.
I loaded $10 in credit and expected it to last at least a week, possibly even a month. After one day of moderate use, nearly all of it was gone.
Agent usage adds up differently from a normal chatbot conversation. A single task can involve multiple model requests as the agent calls tools, controls a browser, runs terminal commands, evaluates results, and decides what to do next.
By the end of the day, Hermes had made 167 requests to Claude Sonnet and processed roughly 16 million tokens. Sonnet accounted for $8.97 of my $9 total spend.
I still like Sonnet, but at that rate it is too expensive for me to use as the main model in an always-available agent.
I have switched Hermes to DeepSeek V4 Pro. It costs significantly less and is designed for reasoning, coding, and agent workflows.
Here’s to hoping DeepSeek produces output of similar quality for a fraction of the cost.
