Feature request: provider batch APIs (Anthropic Message Batches, OpenAI Batch, etc.)

[husseinraoouf]

The gap

Several providers TanStack AI already supports have a batch API: you submit many requests at once and collect the results later, at a lower price. Anthropic's Message Batches, for example, cost 50% of the standard price and return results within 24 hours. OpenAI, Gemini and Bedrock each offer a similar batch mode.

As far as I can tell, TanStack AI has no way to use these today. Searching the repo turns up nothing for messages.batches or batches.create, and the batch mentions in the docs are about stream chunking and Bedrock embeddings. So an app that wants batch pricing has to drop down to each provider's own SDK for that one path. It then loses what it uses TanStack AI for: the same request shape across providers, typed outputSchema results, image input, and provider switching.

Our use case

We use TanStack AI with Claude (via @tanstack/ai-anthropic) to turn messy shop data, such as social-media posts with images and captions, into structured products using outputSchema. A large import can mean hundreds of independent requests that don't need to finish in real time. These requests are a natural fit for batch pricing. Today the batch path has to be written separately for each provider, with its own request and result types.

Rough idea

This is only a sketch; the maintainers will know the right shape far better. It's an activity that takes the same request options as chat(), minus streaming, and splits submitting and collecting into two steps, so it fits durable or long-running workflows:

const batch = await submitBatch({
  adapter: anthropicText("claude-opus-5-5"),
  requests: posts.map((post) => ({
    customId: post.id,                 // caller's ID, used to match results
    messages: [/* text + image parts */],
    outputSchema: ProductDraft,
  })),
})
// persist batch.id, then later (possibly in another process):
const status = await getBatch({ adapter, id: batch.id })        // in_progress | ended | ...
for await (const result of batchResults({ adapter, id: batch.id })) {
  result.customId   // matches the request
  result.output     // typed by outputSchema, or a per-request error
  result.usage      // token usage, for cost tracking
}

Useful properties:

• Per-request results. Each request succeeds or fails on its own, with a typed error.
• Same features as chat(). Typed output and usage work the same way for batched requests.
• Cancel. A way to cancel a submitted batch.
• Capability flag. Adapters say whether they support batching, so apps can fall back to running requests at once with a concurrency limit (as the Bedrock embeddings docs describe for Titan).

Happy to give more detail or test an early version with Anthropic.