Completion models are now considered legacy.
When I am in the playground running tests with completion models and can’t help but notice this text on the bottom screen. Does that mean completion mode will be deprecated some time in the future to focus on chat solely? Are we being advised to migrate to chat mode some time down the line?
Welcome to the community!
It’s quite possible. While they’re still releasing new completion models like gpt-3.5-turbo-instruct (which isn’t legacy) - it’s quite possible that in the long run, they might be going away eventually, or simply not be as supported as the chat models 
I think it’s safe to assume that some open source model might become a drop-in replacement for gpt-3.5-turbo-instruct if it becomes deprecated - but I don’t know if we can expect active development of considerably stronger completion models. Maybe cheaper models, but maybe not stronger.
so, tl;dr: gpt-3.5-turbo-instruct isn’t legacy yet. but it looks like the industry is moving away from that 
From OpenAI:
We use the term “legacy” to refer to models and endpoints that will no longer receive updates. We tag endpoints and models as legacy to signal to developers where we are moving as a platform and that they should likely migrate to newer models or endpoints. You can expect that a legacy model or endpoint will be deprecated at some point in the future.
Then despite that clarification, “legacy” and deprecations" continue to be used inconsistently in other documentation, often with imagined meaning.
The completions endpoint at one point in time was “will be deprecated Jan 2024” - meaning shut off. But it is still here with a bit of backpedal and re-categorization into a new term.
Legacy means it will be there for backward compatibility. There’s no allusion to any timeframe. Could be forever.
Deprecation means it will be there for backward compatibility and will eventually be removed. Deprecation does not mean immediate removal.
Both imply “Don’t use these going forward. We have newer, better APIs to perform the same functionality”.
Newer is not always better. ChatML is great in many ways, but for some use cases (like fiction) the old “instruct” non-chat completion models are better. However, they’re harder to keep “safe.”
I’m hoping trusted users might have access to a GPT-4-instruct or even a GPT-4-fiction in the future, but we shall see. (Maybe @AndrewMayne can be our Champion in our time of need? Small smile…)
Seriously, though, it all depends on your use case. It has become more difficult to get GPT-4 to stray from being a chatbot in some ways, but it is what it is, I guess. 
Yep. “Better” is usually the goal from the API provider’s POV. It is not always the case from the developer’s POV.
Case in point… when writing fiction it’s easy for a chat-bot to forget its role as “author” and break the fourth wall haha…
So, yeah, some of us are championing for it, but honestly, I don’t see it happening.
Anthropic and others might be only choice? Or that style of completion is gone forever? It works well for some tasks, but it’s easily misled which is ripe for abuse… Which brings me back to a “trusted” user status of some sort at OpenAI…
