Economics of custom models
Worked numbers on what custom models cost to train and serve, and when they pay for themselves.
Articles
Prompted open model vs fine-tuned open model vs frontier API: a quality-adjusted cost model
Price per token is the wrong comparison: once failures carry a cost, accuracy and volume decide whether a prompted open model, a fine-tune or a frontier API is cheapest.
Read →Economics of custom modelsCost per million tokens, from first principles
Cost per million tokens is GPU price divided by useful tokens per paid GPU-hour; batch size, utilisation and the input/output mix decide almost all of it.
Read →Economics of custom modelsCost per task, not per token: retries, tools, long contexts and human escalation
For agentic tasks, tokens are usually a small share of cost: success rate, verifier recall, escalations and silent failures decide what a task really costs.
Read →Economics of custom modelsThe ROI of post-training: when a fine-tune pays back
A fine-tune is an investment: its payback depends mostly on volume, the cost of failures and whether it measurably beats the baseline, so a cheap pilot is often the best first spend.
Read →Economics of custom modelsSelf-hosting vs API: the total cost of ownership of LLM serving
Below billions of tokens a month, self-hosting cost is set by engineering time and a redundancy floor, not by GPU price or throughput.
Read →Economics of custom modelsUtilisation and batching economics: why idle GPUs dominate the cost of a token
Cost per token is hyperbolic in utilisation, and with continuous batching traffic sets the batch, so a lightly loaded GPU pays for idleness twice.
Read →Bring a model and a metric.
Tell us what the model should get better at. A post-training engineer replies within one business day with a plan, a compute estimate and an eval design.