The 50% batch discount is the biggest published lever on a Claude bill, so batch support should weigh into platform selection for any workload with an asynchronous component. The confusion to clear up first: Anthropic's Message Batches API (the /v1/messages/batches endpoint) is one specific product. AWS and Google Cloud each offer their own cloud-native batch inference for Claude — different mechanisms, same headline discount. Saying "Bedrock has no batch" is wrong; saying "Bedrock has Anthropic's Batch API" is also wrong.
The map, as of July 2026
| Platform | Batch mechanism | Pricing | Key constraints |
|---|---|---|---|
| Claude Platform on AWS | Anthropic Message Batches API | 50% off standard | Same as first-party; IAM action CreateBatchInference |
| Amazon Bedrock | AWS batch inference (S3-based) | 50% below on-demand | No tool calling or structured output; bedrock-runtime only |
| Google Vertex AI | Google batch prediction (BigQuery / Cloud Storage) | 50% of on-demand | Default 4 concurrent batch requests per project; no global endpoint |
| Microsoft Foundry | None | — | Message Batches API explicitly not supported |
Claude Platform on AWS: the full Anthropic Batch API
Because Claude Platform on AWS is Anthropic-operated and exposes the Claude API surface directly, batch processing works as documented for the first-party API: up to 100,000 requests or 256 MB per batch, most batches done in under an hour, a 24-hour expiry after which unprocessed requests are not billed, results downloadable for 29 days, and support for vision, tool use, system prompts, and extended thinking inside batched requests. Access is governed by AWS IAM — CreateBatchInference, GetBatchInference, ListBatchInferences, CancelBatchInference, and DeleteBatchInference actions on the workspace ARN — which also means you can deny batch to specific principals, useful because the Batch API is not eligible for zero-data-retention arrangements.
Amazon Bedrock: AWS-native batch inference
Bedrock does not offer Anthropic's Message Batches endpoint. Instead, AWS provides batch inference on the bedrock-runtime side: you upload inputs in InvokeModel or Converse format to S3, and outputs are delivered back to S3, at "a 50% lower price compared to on-demand inference pricing" per AWS. Constraints from AWS's documentation: it is not supported for provisioned models, and it does not support tool calling or structured output (response_format) — a significant functional gap versus Anthropic's batch, which supports all server tools. The IAM action is bedrock:CreateModelInvocationJob, and batch quotas apply only to bedrock-runtime, not the newer bedrock-mantle endpoint.
Google Vertex AI: batch prediction
Vertex likewise lacks the Anthropic endpoint but has Google-native batch prediction for Claude: input from a BigQuery table or JSONL files in Cloud Storage, with "Batch Input"/"Batch Output" rates listed at 50% of on-demand on Google's pricing page (Opus 4.8 batch: $2.50 in / $12.50 out per MTok). Documented constraints: a default of 4 concurrent batch requests per project, and the global endpoint is not supported for batch — plan for regional capacity and note that regional endpoints carry a 10% premium over global for Sonnet 4.5-and-later models on this platform.
Microsoft Foundry: no batch today
Anthropic's Foundry documentation lists the Message Batches API as not supported at all — on either the Hosted-on-Azure or Hosted-on-Anthropic deployment option — and Microsoft documents no Azure-native batch alternative for Claude. High-volume asynchronous Claude workloads on Azure today run at standard CCU-metered rates, or route to another platform for the batch leg. Multi-platform failover is explicitly supported by the ecosystem's design (Claude Platform on AWS, for instance, uses a separate capacity pool precisely so workloads can span platforms).
Where to go next
Read streaming vs. batch cost profiles for which workloads to move, and the Bedrock batch workaround guide for S3-based specifics. The platform overview frames the bigger selection question.