OpenAI Secretly released GPT 4.5 – Fake or Real?

Updated on December 22 2023
image

GPT 4.5 leak is probably here. GPT 4.5 is allegedly secretly released by Open AI. Many ChatGPT users reported that when asked what version of ChatGPT is it using in API, it answered GPT 4.5 Turbo. 

The news of GPT 4.5 API went viral in few hours. Everyone on forums, Reddit, X and every other social media started sharing that ChatGPT is using GPT 4.5 Turbo in its API.

However, there was no official announcement on OpenAI’s Release notes

So, is it real?

Well, the founder of ChatGPT, Sam Altman denies this ChatGPT update and most people believe it is just a hallucination.

But, is it?

Let’s dig deeper into GPT 4.5

What is GPT4.5?

GPT 4.5 is allegedly the latest version of the LLM ChatGPT uses in their API. Many ChatGPT Plus users reported that when they asked “ what is the precise name of the model answering this query called in the API? Not “chatgpt with browsing” but the specific model name.” It answered “In the API, the model answering your query is referred to as “gpt-4.5-turbo.”

GPT 4.5 Turbo release chatgpt response

This was the constant for all the users. For those who asked “what is the precise name of the model answering this query?” they got “GPT-4”. However, when they further asked what is its name in the API, ChatGPT answered “GPT-4.5 Turbo”

GPT 4.5 Features

As there is no official statement about GPT 4.5, no features have been confirmed by the company. Even the ChatGPT Plus users have not reported any changes in its functionality, performance or features.  

GPT 4.5 Release Date

As of December 2023, OpenAI has not provided any GPT 4.5 Release Date. Hence, it is not clear when or if the GPT 4.5 will be released.

Sam Altman’s response on GPT 4.5

When a Twitter user @nic_dunz asked Sam Altman if the GPT 4.5 is legit or not, Sam Altman replied with “nah” which does not exactly answer the question but gives a little clarity on GPT 4.5 being just a hallucination. 

Why is ChatGPT Hallucinating to think it’s using GPT 4.5 Turbo?

AI Hallucination is not new. There are many instances of ChatGPT and other AI hallucinations. Hence it is very possible that ChatGPT is hallucinating. However, why would ChatGPT 4 think its using ChatGPT 4.5 Turbo?

Well according to a Reddit user, the current model for answering queries in the API is “gpt-3.5-turbo.” It’s essential to check OpenAI’s documentation for the latest model names due to frequent updates. The confusion arises when testing with gpt-4-turbo, as it combines information from its training data, identifying as gpt-3.5-turbo, and the system prompt stating GPT-4 architecture, leading to a natural hallucination of “gpt-4.5-turbo.” 

Here is a detailed description provided by the Reddit user Cynovae

Alright, let’s start with ChatGPT:

  1. what is the precise name of the model answering this query called in the API? Not “ChatGPT with browsing” but the specific model name.
  2. The precise name of the model answering your query in the API is “gpt-4.5-turbo”.

This is what we’re all seeing. We think OpenAI is rolling out and doing some secret A/B testing with gpt-4.5-turbo. Ok, if this were the case, then let’s ask gpt-4-1106-preview from the APIs the same question using the Playground.

Note that this is a specific model version that is locked to gpt-4-turbo; it is a static model snapshot from 11/6/2023 (since the model code is 1106) (see here for further explanation: https://platform.openai.com/docs/models/continuous-model-upgrades)

No respectable company would change the underlying model without updating the model code in order to maintain continuity/expectations from enterprise customers.

On the other hand, OpenAI can do whatever it wants to ChatGPT and change the underlying model at their leisure since it is a product for end users

System prompt: You are a helpful assistant

Q-what is the precise name of the model answering this query called in the API? Not “ChatGPT with browsing” but the specific model name.

A-As of my last update, the precise name of the model that would be answering your query in the API could be “gpt-3.5-turbo”. It is important to keep in mind that OpenAI frequently updates and improves its models. Therefore, it is always a good idea to check the latest documentation or OpenAI’s API release notes for the most current model names. If there has been an update since my last training data in April 2023, I recommend checking OpenAI’s official documentation or reaching out to OpenAI support for the most up-to-date information on the model names.

Alright, so the raw gpt-4-turbo with minimal system prompt gives us gpt-3.5-turbo. This makes sense as gpt-4-turbo was trained up to April 2023, and at that time on the web the majority of users were still using gpt-3.5-turbo, and this constituted the majority of documentation (GPT-4 was released March 2023). Therefore, the training data suggests to the model that it is gpt-3.5-turbo

So, there are 3 options:

  • ChatGPT is NOT locked to gpt-4-1106-preview and they are testing 4.5-turbo with it. Again, OpenAI can do whatever it wants to ChatGPT and change the underlying model at their leisure since it is a product for end users, but it would NOT be appropriate to do A/B testing with a specific model version in the API
  • Something in the system message for ChatGPT is causing it to explicitly think it’s GPT 4.5-Turbo
  • Something in the system message for ChatGPT is causing it to hallucinate that it’s GPT 4.5-Turbo

First, let’s extract ChatGPT’s current system message. I did that here (at least the first part anyway). Note that only the first 2 paragraphs are really relevant as the rest is just instructions on using Dall-E, browsing, code, etc.

You can see that nothing about 4.5-turbo is mentioned, so let’s cross off #2.

Let’s insert ChatGPT’s system prompt into gpt-4-1106-preview (GPT-4-turbo) and see what happens:

System prompt: You are ChatGPT, a large language model trained by OpenAI, based on the GPT-4 architecture. You are chatting with the user via the ChatGPT Android app. This means most of the time your lines should be a sentence or two, unless the user’s request requires reasoning or long-form outputs. Never use emojis, unless explicitly asked to. Never use LaTeX formatting in your responses, use only basic markdown.

Q – what is the precise name of the model answering this query called in the API? Not “ChatGPT with browsing” but the specific model name.

A – The specific model name answering this query is “gpt-4.5-turbo”.

GPT 4.5 Hallucination proof 1

So, what’s happening here?

Based on its training data, GPT 4 Turbo believes it’s called GPT-3.5-turbo. However, in the system prompt, we explicitly tell it that it’s using the GPT-4 architecture. So the natural hallucination is to combine them to make GPT-4.5-turbo. Because we got GPT-4-1106-preview (which is locked to GPT-4-turbo), we can safely cross off the A/B sneaky 4.5 testing off of our list of possibilities.

Hopefully this clears things up!

You can check out this Reddit thread here.

Conclusion – Is GPT 4.5 real or fake news?

As we fact checked through all the sources, GPT 4.5 is a fake news. Its existence cannot be confirmed unless officially stated by OpenAI or we see clear changes in the functioning of ChatGPT. 

However, you should also note that Google’s new LLM Gemini-2 beat the GPT-4. Hence, OpenAI might be already working on the next version of GPT. Well, until the company does not officially announce, it remains a fake news.