Unfortunately, that’s not a true characterization, as OpenAI has treated the chat models differently as time went on.
Upon introduction of the chat endpoint, there was a gpt-3.5-turbo model, and a gpt-3.5-turbo-0301. The latter with the date was to be a snapshot, while the gpt-3.5-turbo indeed get constant changes, aggressively so through April and May.
In June a new model was announced, and a new scheme where gpt-3.5-turbo would be an alias pointed to the currently-recommended model. With the creation of gpt-3.5-turbo-0613, gpt-3.5-turbo was then (and is still is) pointed to that model two weeks later, with the existing “real” model gpt-3.5-turbo, that had accumulated changed, being turned off and no longer accessible.
gpt-3.5-turbo-0613 however is NOT a snapshot. It has continued to accumulate undocumented changes to its training and performance, oftentimes breaking applications. There was no alternative for those seeking reliability with the capability of calling functions that was introduced with this model.
gpt-3.5-turbo-1106 is the “preview/beta” model introduced at devday, with 16k input context length at no greater cost. It has been trained on also using parallel tools. However, it seems, due to issues, it is no longer on the path to being a gpt-3.5-turbo replacement in its current form (which was slated for December 11), with now the requirement for a new gpt-3.5-turbo version under development and due next year.
So currently: gpt-3.5-turbo → gpt-3.5-turbo-0613 → continued changes since introduction.
GPT-4 has followed a similar path. gpt-3.5-turbo-0301 also proved not immune to OpenAI tampering.
The API now offers a system_fingerprint that is supposed to report back these previously undocumented alterations that affect performance and determinism.