Hello Google Gemini API team,
An existing image batch is reported as completed successfully, but none of its 334 image results can be retrieved. Please investigate and help recover the existing output.
Batch: batches/hgjz2fqulkvh0ef9ixve4s8xofzcit1755xj
API: Gemini Developer API, generativelanguage.googleapis.com, v1beta (not Vertex AI)
Submitted model reference: gemini-nano-banana-2.1
Created: 2026-10-15T08:38:49.037175992Z
Reported end time: 2026-10-15T08:48:49.158007641Z
Request count: 334; inline batch input; imageSize 2K; aspectRatio 16:9
Observed behavior:
- batches.list reports done=true and BATCH_STATE_SUCCEEDED.
- GET /v1beta/batches/hgjz2fqulkvh0ef9ixve4s8xofzcit1755xj repeatedly fails with 503 UNAVAILABLE or 504 DEADLINE_EXCEEDED, or returns HTTP 200 containing only name and done=true.
- The HTTP 200 responses have no metadata.output, response, inline image results, or output-file reference.
- Files.list exposed the uploaded reference files but no result file for this batch.
- Three other batches using the same configured API key returned 105 images in total, which were downloaded and decoded successfully. A smaller reference-batch GET control returned inline image data with the same diagnostic headers.
Recorded examples (all times UTC, 2026-10-15):
11:02:47.462865 — GET returned HTTP 504 after 60.33 seconds:
{“error”:{“code”:504,“message”:“Deadline expired before operation could complete.”,“status”:“DEADLINE_EXCEEDED”}}
11:09:58.920733 — standard GET returned HTTP 503 after 60.30 seconds:
{“error”:{“code”:503,“message”:“The service is currently unavailable.”,“status”:“UNAVAILABLE”}}
11:10:59.226475 — GET with X-Server-Timeout: 180 returned HTTP 200 after 59.68 seconds:
{“name”:“batches/hgjz2fqulkvh0ef9ixve4s8xofzcit1755xj”,“done”:true}
This was the complete decoded response (77 bytes), not a filtered status display.
Already tried: repeated read-only retrievals with delays, increased server timeout (including 600 seconds), gzip, explicit fields/field-mask headers, and a smaller successful control batch. No workaround has returned the affected image results. The root cause has not been established; a response-size issue is only a hypothesis.
Expected: retrieve the existing per-request results or a downloadable result file.
Actual: 0/334 image results retrieved from the affected batch.
Please:
- Investigate why successful completion is reported but no output can be retrieved.
- Recover/export the existing results to downloadable JSONL or provide another supported retrieval method, if possible.
- Confirm whether these 334 generation requests were charged and explain how to request a billing adjustment if the output cannot be recovered. I have not verified whether this batch was billed.
The batch has not been cancelled, deleted, or resubmitted. I want to avoid regeneration and duplicate charges.
I can provide project/account identifiers through an official private support channel if needed.
Related report (similar symptoms; not proof of the same root cause):
Sanitized retrieval logs are available. Thank you.