{"slug":"self-hosted-models-cost-wall","title":"Self-hosted models on real planning tasks: a cost wall","kind":"Method","blurb":"Where open models stop being cheaper. Benchmarks and the spreadsheet behind them.","published":"2025-11-22","board":"reasoning-on-self-hosted-models","cite_as":"Xooplab (2025). \"Self-hosted models on real planning tasks: a cost wall.\" xooplab.com/publications/self-hosted-models-cost-wall","read_minutes":5,"claims":[{"id":"self-hosted-models-cost-wall#claim-1","text":"On a 14-task financial-planning benchmark, the best open ≤14B model reaches 71% of frontier accuracy at ~6% of the cost."},{"id":"self-hosted-models-cost-wall#claim-2","text":"Above ~78% accuracy target, open self-hosted models stop being the cost-effective choice on commodity GPUs."},{"id":"self-hosted-models-cost-wall#claim-3","text":"Thin instruction-tuning on the benchmark's task family is dramatically cheaper than heavier prompting."},{"id":"self-hosted-models-cost-wall#claim-4","text":"Inference cost dominates total cost only above ~10k requests/month; below that, dev + ops dominates and the calculus flips."}],"url":"https://xooplab.com/publications/self-hosted-models-cost-wall"}