Content
summary Summary

It seems that GPT-4.5 Turbo has been leaked by the OpenAI web team. The first search engines like Bing and DuckDuck Go indexed the GPT-4.5 Turbo product page before an official announcement.

Ad

However, the indexed link for GPT-4.5 Turbo leads to a 404 page. But there is some teaser text visible in the search engine that says GPT-4.5 Turbo is OpenAI's fastest, most accurate, and most scalable model to date. What exactly is behind these terms remains to be seen.

Image: Matthew via X

A likely date for the official unveiling is this coming Thursday - the anniversary of the release of GPT-4. OpenAI CEO Sam Altman has also signed up for Lex Fridman's podcast. There was a Fridman podcast with Altman after the release of GPT-4.

GPT-4.5 Turbo with larger context window and up-to-date information

The context window, which determines how many words the model can process simultaneously, is specified in the teaser text as 256,000 tokens - twice as many as GPT-4 Turbo's 128K. 256,000 tokens is about 200,000 words.

Ad
Ad

More important than the amount of input data, however, is how reliably the model processes all that data. Current models with large context windows tend to ignore information, making them less valuable for text analysis of large documents, for example.

If GPT-4.5 Turbo could perform better here, the model would be a big win, even if the performance is otherwise similar to the current models. Google has a powerful large context model in beta testing with Gemini 1.5.

The teaser description also indicates that GPT-4.5 Turbo is up-to-date until June 2024, the so-called "Knowledge Cutoff Date". This suggests that the new Turbo model will be released in June. This would be rather atypical for OpenAI: Up to now, newly introduced models have been available immediately after their launch.

It could be a strategic maneuver. OpenAI's competitor Anthropic recently released Claude 3, a model that is at least on par with GPT-4 in many disciplines. OpenAI might want to regain model leadership as soon as possible to prevent customers from switching to Anthropic, even if it's only on paper until June.

There have been rumors about an upcoming GPT-4.5 Turbo release since at least December 2023. The rumors say that the new model could have video or 3D capabilities in addition to text and images. The teaser leak doesn't mention multimodal capabilities.

Recommendation

A few days ago, OpenAI CEO Altman held out the prospect of a new AI model "worth the wait."

Ad
Ad
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.
Support our independent, free-access reporting. Any contribution helps and secures our future. Support now:
Bank transfer
Summary
  • It seems that OpenAI GPT-4.5 Turbo has been leaked: search engines like Bing and DuckDuck Go have indexed the product page. The link still leads to an error page.
  • The model is said to have a context window of 256K tokens, twice as much as GPT-4 Turbo, and will be up-to-date until June 2024. It is also said to be OpenAI's fastest, most accurate, and most scalable model to date.
  • There have been rumors about GPT-4.5 Turbo since December. The leaked teaser text does not provide any information about possible video or 3D capabilities.
Online journalist Matthias is the co-founder and publisher of THE DECODER. He believes that artificial intelligence will fundamentally change the relationship between humans and computers.
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.