In the Open AI Playground, under Assistants, where you set the model for an Assistant, how can I ensure my assistant is always using the latest model of GPT4o?
Right now, I have my assistants set to GPT4o, as the image below. They’re not set to, say, one of the more recent date labelled models such as GPT40-2024-11-20.
Does keeping my assistant at the default GPT40 ensure I’m using the latest version of the model? If not, how can I ensure my assistants are using the latest versions, short of manually going into each assistant and updating the models each time for each assistant?
Note: I use the assistants via no-code integrations such as Zapier and n8n, so will be unable to run any code myself.
The short name is an alias, a pointer. It is described as a “currently recommended” AI model.
Using gpt-4o currently runs your API request against the model, gpt-4o-2024-08-06. Besides structured outputs, it is based on instruction hierarchy, giving low priority of instruction-following to roles such as tool returns, prior assistant responses, below user inputs, below system instructions, then below OpenAI’s own training.
gpt-4o-2024-11-20 was rather quietly released at the same pricing. It has no new features exposed. What hints we have about it is that it is described as more “creative” or “conversational”, and that it does write more exhaustively.
gpt-4o-2024-05-13, the initial release of the series, does not have the features of accepting and enforcing structured response_format as an API parameter. It costs a bit more to use, and can produce considerably different results, often reversion-worthy.
The transition of switching the pointer to 08-06 took over a month, but seemed inevitable once there were no breaking bugs like gpt-4-turbo-1106 had. I expect that the newer model is also viewed as advantageous and implementable.
OpenAI continues to alter “snapshot” models of the past, moreso than Azure deployments, so a recommendation to build your application on a specific version doesn’t make its operation absolutely stable, but can let YOU decide when to make the larger version switch.
Under documentation (linked on the left of the forum) would be where you can find a categorical listing of AI models.
There is also a release notes section.
There are blog announcements, the announcements section here, and other places for discovery, but these generally cover a generation of AI model, often with pre-release information that doesn’t relate to the dated version that released on the API.
The very first use of a dated version model name along with a short name version was gpt-3.5-turbo-0301 (2023). It was supposed to be that the dated version was a production version of high reliability, with monthly versions to follow, while the “gpt-3.5-turbo” version was the leading-edge one getting regular updates in supervised training. That paradigm and promise was broken within months, then with no stable version at -0613, but rather the pointer and the dated versions being meaningfully impacted by stealth alterations, and even gpt-3.5-turbo-0301 being later damaged.
