Frugon shows which AI calls could move to cheaper models
Frugon is a free, open-source tool that analyzes AI model call logs on your own machine and estimates where cheaper models could do the job. Its creator built it after heavy AI use caused token usage to rise so fast that a could run out before the week ended. Tracking cost by task type showed that a large share of spending went to searches, scans, and simple scouting work rather than .
Frugon reads OpenAI-style logs, calculates current cost, compares that spend against other models, and recommends which calls could be routed to cheaper models while keeping harder calls on the original model. If logs are not already available, a local middle layer can record calls while passing them through unchanged; existing logging systems can also write one call per line for Frugon to read. The basic analysis without , while the optional `--measure` mode sends a sample of prompts to candidate models using the user’s own provider keys so outputs can be compared side by side.
`--judge` lets another model score those comparisons, and the estimates use LiteLLM pricing, LMArena quality tiers, and general savings bands from RouteLLM research. A bundled demo analyzes 56,100 calls and shows monthly spend falling from $549.46 to $343.91, a 37.4% reduction, by moving easy calls to a cheaper model and keeping hard calls on the stronger baseline model.
Key points
- Frugon analyzes real AI model call logs locally and estimates possible cost savings.
- It targets easy or repetitive calls, such as searches and scans, that may not need an expensive model.
- The basic cost analysis does not send data over the network; optional s use the user’s own provider keys.
- The demo shows a 37.4% monthly cost reduction on 56,100 sample calls.
- It is an open-source tool focused on analysis and routing recommendations, not a full automatic routing platform.