Hi everyone,
I’m reporting what appears to be a billing bug with the Batch API when using gpt-5.4-pro.
What happened
On April 16-17, 2026, I submitted several batch jobs using gpt-5.4-pro via /v1/responses endpoint. All batches either:
- Got stuck at 0/10 completed for 7+ hours, then cancelled
- Were cancelled after making no progress
- Failed immediately (reasoning_effort=low not supported by Pro)
The billing issue
Despite all batches returning:
- completed: 0
- output_file_id: null
- usage.output_tokens: 0
The Cost export shows:
- batch api | gpt-5.4-pro, input: $26.41
- batch api | gpt-5.4-pro, output: $185.99
- Total: $212.40 charged for zero delivered output
Activity data discrepancy
The Activity export reveals:
- 98 requests processed internally, but I only submitted 34
- 2,066,526 output tokens generated but never returned
- It appears OpenAI internally retried/reprocessed requests ~3x without my knowledge, charged for all of them, and delivered
nothing
Additional context
- Standard gpt-5.4 batch was also abnormally slow on the same day (stuck at 0/10 for 50+ minutes), but billing was correct
for standard — only charged for completed work - The billing discrepancy appears to be specific to gpt-5.4-pro batch
- I’ve seen the “Batch API degraded since March 4” thread — my issue is related but more severe because of the incorrect
billing - I have filed a support ticket and it has been escalated to a specialist
Questions for the community
- Has anyone else experienced billing for gpt-5.4-pro batch jobs that returned no results?
- Has anyone successfully received a refund for similar batch billing issues?
- Is the March batch degradation issue truly resolved?
Any insights appreciated.