# Azure OpenAI Fine-tuning Job

> Fine-tuning an Azure OpenAI model bills per training token at job completion, with $0 standing cost while it queues or runs. The recurring expense arrives afterwards, as an hourly hosting charge on the deployed fine-tuned model. ZopNight links each job to its resulting model so both halves of the cost stay visible.

Source: https://zop.dev/integrations/azure/azure-openai-finetuning-job
Updated: 2026-08-19

---

Fine-tuning jobs customize base OpenAI models on your data, billed per training token, and resulting fine-tuned deployments carry an hourly hosting charge. Forgotten fine-tuned model hosting is a recurring surprise line item.

## Training tokens now, hosting charges forever after

A fine-tuning job has two distinct cost phases, and only the first is obvious. Phase one is training: billed per training token, settled when the job completes, with no standing cost while the job queues or runs. Phase two begins when the resulting model is deployed: a fine-tuned deployment carries an hourly hosting charge for as long as it exists, independent of whether anything calls it. The training bill is a one-time event a team plans for; the hosting bill is open-ended and easily forgotten, which is why the second phase routinely ends up costing more than the first.

## Connecting a job to the model it produced

ZopNight's Azure OpenAI enricher enumerates fine-tuning jobs from each account's data plane, capturing job status, created and finished timestamps, and, critically, the linkage to the resulting model. That linkage is what turns a transient training record into cost visibility: training spend attributes to the job, downstream hosting attributes to the deployment it produced, and the two can be read together. Fine-tuning jobs are discovery-only; a completed job cannot be stopped or scheduled, and the actionable object is the hosted model it left behind.

## The customization graveyard pattern

The classic leak: a team fine-tunes several candidate variants, ships one, and leaves the runners-up deployed. Each also-ran hosts at an hourly rate with zero traffic. A close cousin is the superseded generation: a fine-tune of last year's base model still hosted after the workload moved to a newer base. Both are invisible in aggregate token spend and obvious the moment jobs and their models are listed side by side.

## Reviewing fine-tuning history

Azure AI Foundry portal → Fine-tuning shows each job with status and output model; from the Azure portal, enter through the Azure OpenAI account hosting the jobs.