Which OpenAI AI models can be used?

configuration
openai
retag add tags

I've got an email from OpenAI that they will cease to provide access to Assistant API v1 on December 18.

I checked the project to which the API key used by Wing was created in and found dozens of assistants created by Wing. The funny thing is that they were all set to gpt-4-turbo, even though when configuring I set the model to gpt-4o-mini and the console did not complain about the model. Now, gpt-4-turbo is not the most expensive model regarding the input/output token price (it's $10 / $30 per 1M tokens for input/output respectively) when compared to eg. gpt-4 ($30 / $60 ) but is a bit pricier than gpt-4o ($2.5 / $10) and o1-mini ($3 / $12) and considerably more expensive than gpt-4o-mini ($0.15 / $0.6), for example.

Also, when asking the AI to refactor, I see increase in API requests and token usage, but strangely, it matches neither the model picked in Wing AI config nor the model set in the Assistant created by Wing. Last, I tried setting the model in Wing to gpt-4o-mini and now refactoring increases the usage of GPT-4o-mini-2024-07-18 (according to openai's activity dashboard). And when using the chat, it does not seem to result in any API requests at all.

What is happening here? Does Wing create an Assistant that uses the gpt-4-turbo instead of the model given in the configuration but then doesn't use it anyway? Also, why create so many? I tried changing the model used by the assistant to see if it changes anything but it immediately broke the AI in Wing. I had to delete the assistant, disable AI in Wing, and then restart it.

Which OpenAI AI models can be used when configuring the AI?

Perhaps having a dropdown with the possible models would be less error-prone than trying to guess the correct name.

gemisigo

asked 2024-12-13 09:12:11 -0600

edit flag offensive 0 remove flag close merge delete

Comments

add a comment see more comments

1 Answer

Sort by

oldest

newest

most voted

AI Refactor and AI Suggest (the auto-completion feature) both should use the model you select in the AI tool's AI provider configuration. AI Chat ignores that and hardwires an older gpt-4 model because the API it uses changed and doesn't support newer models. This needs to be rewritten. I wasn't aware of the December 18th date and am not sure whether it affects us since I remember the date as being some time later in 2025. We'll look into that in more detail.

Wing should be removing assistants that are not in use. Are you running multiple instances of Wing concurrently, or Wing on multiple machines? They won't all go away until those are all quit. But it's certainly possible something is failing to complete the removal (network connections breaking, etc). We still try to clear them later but if it's failing then you can force removal by selecting menu item Edit > Command by Name and entering commend internal-ai-delete-resources. You will need to restart all instances of Wing that were using the deleted resources before AI Chat will work again.

We recommend using gpt-4o as the model. However, you'll likely need to add money to your OpenAI account so you've contributed at least $50 over time, and may need to wait 7 days to get the necessary rate limit. This waiting time is a newish thing and very annoying. gpt-4o-mini can be used in the interim but doesn't seem to produce results of the same quality.

We've avoided adding a drop-down here or doing more than setting a new default model over time since OpenAI is constantly changing. We'll look into querying what's available and adding a drop-down.

Sorry for the confusion. Obviously, anything using AI is new and in constant flux. Always be sure you have the latest Wing by using Check for Updates from the Help menu. Also, don't hesitate to ask us if you have other questions, or you can email support@wingware.com.

4.2k

Wingware Support

answered 2024-12-13 09:32:21 -0600

edit flag offensive 0 remove flag delete link

Comments

I don't want to sound a miser, but there are considerable differences in price for different models. You say gpt-4o-mini does not produce the same quality results and I agree. I also didn't find gpt-4o to be good at solving even my moderately complex problems adequately. It does some horribly stupid things in SQL and I don't think it fares any better in Python. Except, I recognize BS in SQL, but I'm not an expert in Python. I might not notice when it tries and leads me astray. My worries might be unfounded, but I simply do not trust in enough. Yet.

On the other hand, when decomposing the issue into small enough chunks, I deem the answers given by gpt-4o-mini... well... acceptable at least. Since I ask a lot, being able to ask more than 10 times as many of these "small chunk" questions for the same price ... (more)

gemisigo (2024-12-14 03:44:46 -0600) edit

I've had both astonishingly good results with gpt-4o and also very annoyingly useless ones, where I just can't get it to do what I want no matter how I frame the request. However overall I found things improved a fair amount in the move to gpt-4o from earlier gpt-4 models, hence my recommendation to use it. I've used gpt-4o-mini much less, so am not as familiar with how that compares. Yes, there is a huge price difference, but I find my usage of gpt-4o is amounting to mostly less than a dollar a day so it's not really an issue relative to wanting to get the best results possible. Obviously, my tasks are not going to be the same as yours and so our experiences will differ, both in quality of results and costs.

We will try to improve this, so it's easier to switch models and compare the results ... (more)

Wingware Support (2024-12-14 18:50:37 -0600) edit

add a comment see more comments

Your Answer

Please start posting anonymously - your entry will be published after you log in or create a new account. This space is reserved only for answers. If you would like to engage in a discussion, please instead post a comment under the question or an answer that you would like to discuss.

Add Answer

Which OpenAI AI models can be used? edit

Comments

1 Answer

Comments