fix: implement batch generation for OpenAI models

[ianliuy]

What's broken?

Calling the generator with a list of prompts (vectorized/batched calls) crashes for OpenAI models. For example:

model = outlines.models.openai("gpt-4o-mini")
generator = outlines.generate.json(model, MyModel)
results = generator(["prompt1", "prompt2"])  # crashes

Who is affected?

Any user trying to batch multiple prompts with OpenAI (or AsyncOpenAI) models. Single-prompt calls work fine. This was originally reported against v0.1.13, and while the codebase has been completely restructured for v1.0, the underlying user need (batched generation) is still unmet — generate_batch() raises NotImplementedError.

When does it trigger?

Every time model.batch([...]) or generator.batch([...]) is called with an OpenAI model. 100% reproducible.

Where is the bug?

• outlines/models/openai.py lines 295-303: OpenAI.generate_batch() raises NotImplementedError
• outlines/models/openai.py lines 445-453: AsyncOpenAI.generate_batch() raises NotImplementedError

Why does it happen?

When batch support was added to the model hierarchy (commit c350ddc9), models with native batch APIs (Transformers, vLLM) got real implementations, while API-based models (OpenAI, Anthropic, etc.) got NotImplementedError stubs since their APIs don't support multi-prompt-in-one-call.

However, batch can be trivially implemented for API models by looping over individual generate() calls — each prompt becomes a separate API request. The async variant can use asyncio.gather() for concurrent execution.

How did we fix it?

Replaced the NotImplementedError stubs in OpenAI.generate_batch() and AsyncOpenAI.generate_batch() with implementations that:

• Sync: Loop over inputs, calling self.generate() for each prompt
• Async: Use asyncio.gather() to fire all prompts concurrently

This is a minimal, surgical change (~10 lines per method) that follows the existing generate() contract — input formatting, output type handling, and error handling are all delegated to the already-tested generate() method.

The same pattern could be applied to other API models (Anthropic, Gemini, Mistral, etc.) as a follow-up.

How do we know it works?

• Updated test_openai_batch and test_openai_async_batch from asserting NotImplementedError to verifying correct batch results using mock clients
• Tests cover both sync and async paths, including duplicate-prompt edge case
• All 10 non-API-call OpenAI tests pass (no regressions)
• Pre-commit checks pass (mypy, ruff, merge conflicts, debug statements)

cc @RobinPicard @cpfiffer

[RobinPicard]

Thanks for opening a PR! Just 2 things to change on top of the comment below:

• add type hints and docstrings for the methods that are now implemented thanks to the PR
• update the the documentation for the OpenAI model, specifying in particular that it's not true batching

[RobinPicard]

I think return_exceptions=True would make sense here as users may want to access results for threads that did not raise an exception.