Provide access to old (simpler) LLMs

lelovsky

Not really a feature request: It's as simple as not removing access to older language models from the Assistant.

I wrote my custom instructions to work with e.g. ChatGPT 4 nano, the fastest available one. I got used to the way I interacted with it. It provided the desired output 90% of the time. Once the new reasoning models launched, I had been unable to get consistent outputs and, the most frustrating thing, is that they take ages to "reason" an inconsistent output, while also being very expensive and eco-unfriendly.

I am not going to pay premium price to access a cheaper model. Please, consider paywalling only the premium models, while not removing access to what already has been working fine. Switching my custom assistants from what worked to "Kagi Quick" is just wrong in my opinion. I can't get used to other models!

I don't care about reasoning models for getting one line linux commands and writing powershell scripts. I want the fast and models back! Please!

Just add a couple of older models to Professional/Family plan. Ultimate is too expensive for my use.

Anonymous12

Have you tried Custom Assistants? You could make adjustments as you see fit and choose which model may be suitable. Perhaps Kagi's own 'Quick' model?

Thibaultmol

@Anonymous12 that's not the problem that @lelovsky is having. Their custom assistant got switched to Quick because the previous one they were using got deprecated by the assistant team.
They want to remain having access to said older models even when assistant team deprecrates them.
Which honestly makes sense, def if you're using it in a business where you've created a custom assistant for your team and have KNOWN BEHAVIOUR with a specific model with your instructions.

Luis

We're introducing changes to improve how we handle model deprecation and its impact on your Custom Assistants. These should go live this week. I believe they will solve your problem: although we'll continue to deprecate older models, we'll ensure smoother transitions going forward.

see https://kagifeedback.org/d/5597-provide-feedback-when-selected-model-is-no-longer-available/8 for more details

lelovsky

Luis Thank you for your reply; I hope I understand you correctly: You plan to make the automatic upgrade to the latest model, follow the previously selected family of models e.g. from chatgpt 4 to chatgpt 5. I don't think it will solve my problem; It would be better this way, but in the end, any kind of upgrade makes the model a little bit different, which interferes with my workflow.

What I mean is that the older models should still be accessible, because (in my understanding) they are both cheaper to Kagi Team and to Kagi Users. From my limited research they are not deprecated in a way which would make Kagi Team limit the access to them: those models are available in Ultimate plan and I can access them using official APIs e.g. for $5 I can use ChatGPT 3.5 turbo, practically endlessly, using my vibecoded WinForms application. My problem is that I prefer using Kagi Assistant - it has nice UI, and I can access it on multiple devices.

Correct me if I'm being wrong here: my point is, that access to the older models is being deprecated by Kagi Team - and I can't find a valid reason for you to do so. If the problem is UI being cluttered, I guess, you could add an "advanced mode" somewhere in the account settings page.

Anonymous12

Is this recent announcement relevant to the discussion? I suppose since they are not retiring those models yet through the API, no. But it's inevitable I suppose.