If you pay for ChatGPT, be it Plus, Team, or Enterprise, GPT-4 Turbo is now even better. In a Thursday post on X (formerly Twitter), OpenAI announced the latest version of its GPT-4 Turbo model is now live for all paid ChatGPT users, complete with some interesting updates.
So, what’s new? It’s all a bit general, but OpenAI says the newest iteration of GPT-4 Turbo is now better at writing, math, logical reasoning, and coding. You can see from a graph in the announcement post that the April 9 version of GPT-4 Turbo has improved in just about all categories than GPT-4 Turbo from Jan. 25 of this year. Some metrics are more dramatic than others: The largest leap looks in GPQA, an AI benchmark concerning scientific datasets, where GPT-Turbo has improved from 39.7% to 49.1%; and the MATH benchmark, which saw an improvement from 64.2% to 72.2%.
Interestingly, the one area January’s GPT-4 Turbo model excels April’s is in the HumanEval test, which tests an LLMs’ ability to generate code: The former scored 88.2%, while the new model scored just 87.6%. A small difference, but an interesting step back that OpenAI doesn’t address.
What OpenAI specifically focuses on, however, is the improved general experience using ChatGPT with the new GPT-Turbo, which is what users like myself will most likely notice most anyway. The company says when you’re writing with ChatGPT, the model’s responses will be “more direct, less verbose, and use more conversational language.” You can see that from the example OpenAI uses in its post: It asks ChatGPT generate an SMS to remind friends to RSVP to a birthday dinner invite. January’s GPT-4 Turbo model generates a wordy response, that while still potentially useful, does appear AI generated in some areas. The latest model, however, generates a shorter, more casual reminder about the RSVP.
Perhaps more important of all, GPT-4 Turbo’s data set goes up to December 2023. The previous version was stuck at April 2023, which means it’s missing a full year’s worth of context. While the new GPT-4 Turbo won’t know anything about 2024, unless you connect it to the internet, it will have about eight months’ worth of current events more than it did before.
