Skip to content
Batchwork
Esc
↑↓navigate↵open⌘Jpreview
On this page

Anthropic

Anthropic Message Batches with inline submission and same-origin streamed results.

Anthropic accepts Message Batches inline rather than through a JSONL upload. Batchwork translates messages and tools into Anthropic-native blocks, polls the batch object, and streams the provider’s JSONL results URL.

At a glance

Property Batchwork behavior
Output modalities Text
Text inputs Messages, tools, images, PDFs/documents, provider file references
Endpoint /v1/messages/batches
Submission Inline {requests: [{custom_id, params}]}
Results JSONL from a provider-returned same-origin URL
Batch metadata Ignored
Webhooks Managed polling only
Credential ANTHROPIC_API_KEY
Default base URL https://api.anthropic.com

Example

from batchwork import BatchRequest, Batchwork

async with Batchwork() as client:
    job = await client.batch(
        model="anthropic/claude-sonnet-4-6",
        requests=[
            BatchRequest(custom_id="a", prompt="Summarize the first report."),
            BatchRequest(custom_id="b", prompt="Summarize the second report."),
        ],
    )

Lifecycle

  1. Batchwork serializes every request into Anthropic params and removes stream.
  2. POST /v1/messages/batches receives all requests inline.
  3. poll() retrieves /v1/messages/batches/{id}.
  4. results() streams JSONL from the returned results_url.
  5. cancel() calls /v1/messages/batches/{id}/cancel.

Batch-level metadata= is not forwarded. Anthropic request-level user metadata remains available through provider options.

Messages and tools

Consecutive user/tool messages and assistant messages are grouped into Anthropic-native content blocks. System messages move to the top-level system field where possible. Tools, tool results, thinking blocks, files, documents, citations, cache control, MCP servers, containers, and structured output receive provider-specific translation.

Model-aware settings

  • A default max_tokens is selected when omitted and capped for recognized models.
  • Enabling thinking without a budget defaults the thinking budget to 1,024 tokens.
  • Thinking can increase max_tokens and remove temperature, top-k, and top-p.
  • Known models that reject sampling do not receive sampling fields.
  • Known compatible models do not receive both temperature and top-p together.

Anthropic provider options include speed, inference geography, fallbacks, cache control, user metadata, MCP/container settings, context management, output effort, thinking, tool controls, document context/citations, and strict-tool behavior where supported.

Media and security

Supported image and PDF URLs can be passed directly to Anthropic. Other media sources are resolved before serialization.

The provider-returned results_url must match the configured Anthropic API scheme, normalized host, and effective port. Embedded credentials, cross-origin hosts, alternate ports, and HTTP downgrade are rejected before fetching. This also applies when using a custom gateway.

Results

Successful items concatenate Anthropic text blocks into result.text, normalize usage, and retain the raw message in result.response. Tool-use and other non-text blocks remain available through the raw response. Provider errored, expired, and canceled items map to normalized result statuses.

Anthropic does not support embedding or image-generation output through Batchwork.

Official references

See Text and messages, Results, and Security.

Was this page helpful?