# Azure OpenAI Batch Job

> Azure OpenAI batch jobs run bulk inference asynchronously at 50 percent of on-demand per-token pricing, with zero standing cost, so spend lands when the job completes. ZopNight enumerates batches through the account's data plane with job status and volume context, confirming bulk workloads actually take the discounted path.

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

---

Batch jobs process large volumes of OpenAI requests asynchronously at discounted per-token rates. They are the cost-efficient path for bulk inference, and visibility confirms teams are actually using it.

## Half-price tokens, settled at completion

A batch job charges per token at 50 percent of the on-demand rate, and it charges nothing while queued or running. The whole cost lands when the job reaches a terminal state. In exchange, results arrive asynchronously rather than interactively. That makes batch the correct meter for evaluation runs, embedding backfills, dataset labeling, and any workload where nobody is waiting on a socket. The economics are simple and one-directional: identical tokens routed through a Standard deployment cost twice as much, so every bulk workload still using synchronous calls is a doubled bill hiding in plain sight.

## Why a transient execution still earns an inventory row

Jobs come and go, so treating them as resources may seem odd, but the history is the signal. ZopNight's Azure OpenAI enricher lists each account's batches from the data plane, recording job identity, status, and creation and terminal timestamps, and folds that volume context into usage attribution. A team whose token spend is high while its batch list is empty is paying interactive rates for offline work; a team whose batches fail repeatedly is paying for retries that never produce output. Neither shows up in Cost Management alone, because the invoice does not say which meter the tokens should have used. Batch jobs are discovery-only: transient executions with nothing to schedule or stop.

## Signals that the discount is being missed

Two patterns recur: nightly pipelines calling Standard deployments in tight loops (bulk-shaped traffic on the interactive meter), and one-off migration or backfill scripts written against the synchronous API because batch was unfamiliar. Both halve their token cost by moving to batches.

## Where batch activity surfaces

Azure AI Foundry portal → the account's Batch jobs view lists submissions with status and timing; the Azure portal reaches the same place through the Azure OpenAI account.