Batch API Not Available on Bedrock or Vertex

The mistake

Designing a cost-saving strategy around the Batch API for a deployment that actually runs through a third-party hosting platform — Amazon Bedrock, Google Vertex AI, or Microsoft Foundry — rather than calling OpenAI directly, and only discovering the mismatch once the integration is built.

Why this happens

A lot of published OpenAI capabilities are described in ways that read as platform-agnostic, and Batch's discount is genuinely appealing enough that it's easy to plan around before checking whether it's actually offered on the specific platform your deployment uses. It isn't offered on any of the three major third-party hosting platforms this site tracks — it's an OpenAI-direct capability specifically, not a feature of "the models" independent of where they're being called from.

Why it matters

A cost model built assuming Batch-level pricing for a Bedrock, Vertex or Foundry deployment will be wrong in a specific, unfavorable direction — the real bill lands at the Standard rate for that platform, not the discounted one the plan assumed, and that gap only becomes visible once real invoices start arriving. For a large-volume, cost-sensitive workload, that's not a rounding error; it can be the difference between a project's cost model holding up and quietly not holding up.

The fix

Check platform availability before designing around Batch, not after. If a workload's cost model genuinely depends on the Batch discount, that's a real reason to run it through OpenAI directly rather than through a hosting platform that doesn't offer it — a decision worth making deliberately, weighed against whatever else the hosting platform is providing (existing cloud billing consolidation, data-residency requirements, existing infrastructure integration), rather than discovering the constraint after the architecture is already committed.

This is worth checking early specifically because the fix isn't a config change once discovered late — a workload committed to a third-party platform for other reasons (existing infrastructure, procurement, compliance) genuinely can't recover the Batch discount without moving the workload itself, which is a much bigger decision than swapping an API parameter.

See also

Where Each Model Is Actually Available covers platform-by-platform differences more broadly — Batch availability is one of several things that don't travel uniformly across every place a model is offered.