I noticed that the floating gpt-4o-realtime-preview tag still points to the older snapshot gpt-4o-realtime-preview-2024-12-17 rather than the newer gpt-4o-realtime-preview-2025-06-03.
This can be verified both in the docs and through my own usage.
Presumably, this delay is intentional to give early adopters time to identify any issues with the new release.
The switch from October to December versions of realtime models was very fast - likely because it fit OpenAI priorities.
Here’s a quote from OpenAI blog 2023: the first time that there was now an underlying model to the shorter name’s alias (then a shift from the short-lived idea of the short name being “continually updated” was quite quick; now it is a pointer that is updated after a delay.*).
We look at a large number of evaluation metrics to determine if a new model should be released. While the majority of metrics have improved, there may be some tasks where the performance gets worse. This is why we allow API users to pin the model version. For example, you can use
gpt-4-0314instead of the genericgpt-4, which points to the latest model version. Each individually pinned model is stable, meaning that we won’t make changes that impact the outputs.
The evaluation metrics for updating the alias to point to a different dated version would likely be…not a bunch of forum reports about choppy audio and non-performance as has been seen.
* or not updated - gpt-4o not even yet pointing to the newest gpt-4o-2024-11-20 of six months ago that cannot be fine-tuned
Interesting! If it can take as long as six months in some cases, I’ll definitely keep an eye on it. I’ll give it a week, and if the floating tag still hasn’t updated, I’ll go ahead and pin the newer snapshot myself.
The old snapshot could really use improvements in instruction following, interruption handling, and tool calling.