[bot] OpenAI: Batch API (`client.batches.create()`) not instrumented

[braintrust-bot]

Summary

The OpenAI Batch API (client.batches.create()) is not instrumented. Calls to submit or retrieve batch inference jobs produce zero Braintrust tracing. This is a GA feature in the OpenAI Python SDK for running batch inference (chat completions, embeddings, responses) at 50% reduced cost.

The Braintrust Anthropic integration does instrument the equivalent batch API (client.messages.batches.create() and client.messages.batches.results()), making this an asymmetry across providers in this repo.

What is missing

OpenAI Resource	Method	Instrumented?
client.chat.completions	create()	Yes
client.responses	create(), parse()	Yes
client.embeddings	create()	Yes
client.batches	create()	No
client.batches	retrieve(), list(), cancel()	No (CRUD — lower priority)

The generative-execution-relevant surface is batches.create() (submitting a batch of model inference requests). The retrieve/list/cancel methods are CRUD management and lower priority.

At minimum, instrumentation for batches.create() should create a span capturing:

• Input: input file ID, target endpoint (e.g. /v1/chat/completions, /v1/embeddings)
• Output: batch ID, processing status, request counts
• Metrics: submission latency
• Metadata: endpoint, completion window, batch metadata

This mirrors the pattern already used by the Anthropic batch instrumentation (Batches and AsyncBatches wrapper classes in py/src/braintrust/integrations/anthropic/tracing.py), which creates task-type spans with input metadata, output status/counts, and model information.

Batch API capabilities

The OpenAI Batch API supports batching requests to:

• /v1/chat/completions
• /v1/embeddings
• /v1/responses

Both sync (Batches) and async (AsyncBatches) client variants exist in the OpenAI SDK. The API is in the main namespace (not beta), indicating GA status.

Braintrust docs status

not_found — The OpenAI integration page documents chat completions, structured outputs, function calling, and streaming. No mention of batch API support.

Upstream sources

• OpenAI Python SDK Batches class: openai/resources/batches.py — defines create(), retrieve(), list(), cancel()
• OpenAI Batch API guide: https://platform.openai.com/docs/guides/batch
• The SDK exposes both Batches and AsyncBatches classes with with_raw_response and with_streaming_response variants

Local files inspected

• py/src/braintrust/integrations/openai/patchers.py — defines patchers for ChatCompletions, Embeddings, Moderations, Audio, Images, Responses; zero references to batch or Batch
• py/src/braintrust/integrations/openai/tracing.py — wrapper functions for chat, embeddings, moderations, audio, images, responses; no batch wrappers
• py/src/braintrust/integrations/openai/integration.py — integration class registers patchers; no BatchPatcher
• py/src/braintrust/integrations/openai/test_openai.py — no batch test cases
• py/src/braintrust/integrations/anthropic/tracing.py — Anthropic batch API IS instrumented (Batches, AsyncBatches classes at lines 565-630) as precedent
• py/noxfile.py — test_openai sessions exist but no batch coverage

Relationship to existing issues

• [BOT ISSUE] Mistral: Batch Jobs API (client.batch.jobs.create()) not instrumented #272 tracks the same gap for Mistral (client.batch.jobs.create())
• The Anthropic integration already instruments client.messages.batches.create() and .results() — establishing the pattern and precedent for batch inference tracing in this repo