OpenAI vs Claude vs Gemini: Batch APIs for Document Processing

Compare file submission, request identifiers, partial failures and result retrieval for offline document workloads.

Reviewed by FluidRWA Research Team · November 15, 2026

Short answer

OpenAI and Gemini document file-based batch workflows, while Claude's Message Batches API accepts an array of individually identified Messages requests. Pick the interface that fits your queue and rerun policy. With all three, the end of a batch is a signal to reconcile individual results, not to assume every input succeeded.

Suppose a compliance team submits a month's worth of disclosures for classification. The work can run overnight, but the team still needs a result or a reason for failure for each document. A cheaper request is not much help if operators cannot tell what went missing. Compare these products on the work your queue has to do after submission.

Side-by-side comparison

ProviderWhat the provider documentsSource
OpenAIUpload a JSONL input file, create a batch against a supported endpoint and inspect output and error files. Requests carry a custom identifier.Official documentation
ClaudeSubmit an array of Messages requests, each with a custom_id. The results record identifies the outcome of each request.Official documentation
GeminiSubmit inline requests or use a JSONL file for larger jobs. Retain record keys and inspect the documented response format for the chosen submission method.Official documentation

OpenAI: What to verify

Use the document ID and processing version as the basis for your request identifiers. Keep the model and endpoint explicit in the manifest. When the batch ends, join output and errors against input by identifier; never assume file order matches the input. Have the on-call operator trace one failed record all the way back to the document revision.

Claude: What to verify

Build from the Messages requests you already use so your single-record test and batch test have the same task definition. Validate request shape locally, then practice handling per-request errors after processing. A successfully created batch does not validate that every request can run. Separate requests that need prompt or input correction from ones that can be retried unchanged.

Gemini: What to verify

Try both forms only if both are candidates for your workload. Inline submission may suit tightly bounded jobs; file-based work is easier to partition and archive for a pipeline that already stages batches. Verify how output is retrieved and how record-level errors appear. Recheck the quotas and models for the project you will actually deploy.

Run the same acceptance test with each option

ScenarioPassing evidence to ask for
Retry after job expiryOriginally successful records are skipped. A second job has its own identifier but retains the original document references.
Malformed input mixed into a valid batchThe manifest shows which document failed and who should fix it, without forcing operators to infer failure from the absence of a result.
Wrong schema despite a successful API outcomeA separate validator blocks incomplete or malformed evidence from being used downstream. Corrected results retain the original model and prompt history.
Results arrive too late for reviewThe team can stop relying on that job for the current deadline and reconcile any late results without publishing the same record twice.

What to include in the buyer brief

Keep the input manifest outside the provider. Give each document revision a stable ID and store the checksums, processing timestamp and prompt version in your system. Track acceptance, processing and validation separately. A response can be technically successful and still extract the wrong reporting period or miss a qualifying sentence. For sensitive materials, have someone confirm batch retention, processing region and whether attached files follow separate lifecycle rules before uploading anything real. Test with synthetic or approved examples first.

Work through a realistic overnight run

Imagine 1,000 source documents due for review at 9 a.m. Include a corrected document that supersedes its earlier version, three deliberately invalid inputs, and at least one valid request whose response fails your output checks. Give every row a job-independent document revision ID. When a second job runs, it should contain only unresolved items. Make the operator demonstrate that the older revision cannot overwrite the corrected one.

Keep raw results in an access-controlled store. Send the review tool only the fields reviewers need, plus a reference to the page or excerpt that supports them. The team should see which rows are ready, which failed technical validation, and which need a human to resolve conflicting evidence. This separation matters if the output is going into a customer record or a compliance decision.

Compare cost per usable result rather than cost per request

Record the number of documents that clear validation by the review cutoff. Include repeat requests and manual investigation in the cost comparison. Track separately what proportion arrived but had to be corrected. Use the same document mix, file-size distribution and review criteria across providers. A sample of short, clean inputs does not tell you how an overnight run handles inconsistent formats.

Ask each provider about account limits and material retention for the exact endpoint and model. Verify any contractual data-handling requirement independently of the batch feature. If a document cannot go to a particular service, route it out before building the input file.

Scope of this comparison

The providers were selected because they offer different documented approaches to this workflow. The table uses official product documentation checked on November 15, 2026. The acceptance tests are FluidRWA's recommendations, not completed tests. Pricing, contractual terms and results in a particular deployment were not verified.

FAQ

Should we move real-time customer requests into batches?

Only if your product can tolerate asynchronous completion and degraded or delayed results. Use the batch for an offline workflow with a separate review deadline, not for a user waiting on the current screen.

Evaluating providers for your project?

Share the workflow and requirements you need to verify, or email contact@fluidrwa.com.

Submit requirementsEmail FluidRWA