Insights on ChatGPT Enterprise Using GPT-4-1106-Preview Based on Context Length Specifications

Community
,
1

Hello everyone,

I’ve noticed that ChatGPT Enterprise was initially released with a specification of 32K context length.

However, the official website now indicates that ChatGPT Enterprise supports a 128K context length, while ChatGPT Plus and ChatGPT for Teams are listed with a 32K context length.

Considering that the only publicly available language model offering a 128K context is the gpt-4-1106-preview, it seems reasonable to deduce that ChatGPT Enterprise is utilizing this model.

contextwindow
contextwindow1334×342 12.5 KB

On the other hand, it seems unlikely that ChatGPT Plus and ChatGPT for Teams are exclusively using a non-preview (turbo) model.

Given this, it would appear reasonable to assume that despite the difference in context length, ChatGPT Plus, ChatGPT for Teams, and ChatGPT Enterprise are all using the same underlying language model.

Taking into account that GPT-4-32K is not the mainstream, my hypothesis seems plausible.

However, I’m interested in hearing your insights or if there’s any official information that could confirm or refute this assumption.

Looking forward to your thoughts and discussions.

4 Likes
2

My understanding is that a lot of models’ context length is very vram bound - so it could very well be the same model, just running on cheaper hardware.

This could potentially also explain why OpenAI API charges by input tokens, and have an output limit; perhaps they delegate to specific nodes with specific hardware configurations by prompt length :thinking:

    flowchart
         GPT-4-->a
         a["tiktoken +4k"] -->8k
a-->16k
a-->24k
a-->32k
a-->...
a-->128k
2 Likes
3

Thank you!
What I really want to know is which version of GPT-4 is the top one.

It’s true that you can’t always rely on what a language model says, but when asked about the knowledge cutoff, it answers April 2023, and it seems to be aware of events up to April 2023, so I’m pretty sure it must be the gpt-4-1106-preview.

I just wanted to be sure.

4

Given that gpt-4-1106-preview (aka gpt-4-turbo) is a reduced-expense model, has the same “lazy” seen in ChatGPT as in direct specification of that model by API, and has been trained on the skills of parallel tool calls required for the retrieval function of agents, it is likely that all ChatGPT is using the latest API model (or a variant even beyond that).

“Context length” thus being then just the input size that the special management endpoint and client software is informed of for conversation and data file length for an account type, by a confused max_tokens specification.

image
image828×476 11.1 KB

(ps, OpenAI, misspelled bismo_settings?)

1 Like
5

Thank you as always!

I noticed that ChatGPT’s GPT-4 is of less quality due to being a reduced-expense model.

In this case, I guess the cutoff of the model’s knowledge is accurate.

I also agree with the view that it’s probably the gpt-4-1106-preview (or a variant of it).

However, it seems like there’s still some confusion about the token count :sweat_smile: