I measure what LLM inference actually costs, down to the GPU.

• qvunex_single.py: one file, no install, no network code. Records cost per finished task, and a call whose cost is unknown is marked unknown, never counted as $0. For self-hosted vLLM, --reconcile checks your client's counts against the server's own counters, and --gpu-rate gives the cost per request at the load you actually ran. Both tested live on vLLM.
• GPU cost data: the same benchmark on T4, L4 and L40S, repeated. On one L40S, minutes apart, CUDA-graph timings repeated within 0.7% on average and eager timings moved up to 78%.
• Reported in pydantic-ai, #8704: responses with no usage were recorded as $0 and slipped past spending limits. Confirmed by their triage.
• Proposed in AgentMeasure, #26: self-hosted servers as a known measurement limit, with the server-side way to close it, tested live.
• Cost audit: I run the meter on a day of your traffic and write up where the money goes.

Contact: qvunexaudit@gmail.com, or open an issue on qvunex.