I have the exact same problem. I have enough balance in my account, and I’m the sole user of my account with a single API key.
I was uploading small jsonl files one by one to 4o, each 50k tokens input and 8.5k tokens in output, well-below the 90k threshold for tier 1. Yet, even though there are no batches running, I keep getting "
Enqueued token limit reached for gpt-4o in organization ###. Limit: 90,000 enqueued tokens. Please try again once some in_progress batches have been completed.".
I get the same error even when I upload the files one by one by hand through the Web UI, the files instantly fail upon uploading. Then after 1 hour, I’m able to upload the same file that previously failed with no issues, but I have to wait for another hour before I can upload the next one.
For reference, I have the max_tokens variable set to 100 per prompt in the file (outputs are always 84 tokens since I’m outputting integer values in a json output).
The problem is, I have about 100 files, and do not have the patience to wait for this bug…
I have the same problem with gpt-4.1. model. It was all working fine untill it didn’t. I made my batches very small and still getting the error about the enqueued token limit even when no batches are running. Was in contact with customer support, but their recommendations have not worked so far.
The same problem with Batches & Embeddings. No batches in progress but any new batch is failing.
I’m thinking about moving to Gemini Embeddings. @OpenAI_Support are you going to fix that?
No, it was a problem related to the incorrect calculation of limits in the Batch API, because when moving to the next Tier, the limits increased, and the problem disappeared
Exactly the same problem here. I saw in another post (to which I cannot link, for some reason) that supposedly @OpenAI_Support had fixed it, but it seems the problems persist….
I eventually fixed the problem, it wasn’t on OpenAI’s side: I was using the wrong version of tiktoken, as so my token estimations weren’t correct. After I did that, my batches were processed properly (so the bug is fixed, indeed).
Hi all,
I keep having the same issue as you describe above. I’m processing chat completions and I’m running many of them.
Usually, the first few batch-jobs work well, but after a while I get the “Enqueued token limit reached for gpt-4o-mini in organization…” error.
However, this is seemingly not due to the size of the batches. I make sure that they are not too enough. I started out with batches of 19 M tokens, but trimmed it down to batches of 1M tokens (my limit is 20M). Then I submit the batches separately, one-by-one, so the queue is definitely never above the token limit.
Weirdly, sometimes they run, sometimes they don’t.
So, my batches should be below the token limit, @j.n.fonseca: What version of tiktoken is ‘the wrong one’? I’m definietly using the latest pip-version and I’m specifying the correct model.
I also think that the JSONL-files are encoded correctly, but how could i actually make sure that they are? Any ideas?
Let me post a few batch IDs, in case a moderator sees this:
a quick update: So it turns out that it must have been the Token Limit after all. The problem disappeared (for now) after upgrading to Tier 4.
I think part of the confusion on my side was because the error message says “Enqueued token limit reached…”, but my queue was empty. Now it seems to me like there is a cool-off implemented in the background. On the one hand you can see this because typically you can still start a 1M-token job after the 10M-job won’t start. On the other hand I saw, that when retrying with similar-sized jobs, the period between the jobs that worked was approx. constant, which would be consistent with the cool down theory.