OpenAI is rolling out a more affordable, slimmed-down version of its flagship artificial intelligence model to appeal to a wider range of developers and business customers in an increasingly crowded market for AI services.
The ChatGPT maker on Thursday announced GPT-4o mini, a smaller, cheaper variation of its most capable AI model, GPT-4o. The San Francisco-based start-up said the updated model will be available immediately for free users and paying ChatGPT Plus and Team subscribers and will be offered to enterprise customers next week. GPT-4o mini is replacing GPT-3.5 Turbo, an older model, in ChatGPT.
OpenAI unveiled GPT-4o in May, showing off its potential as a better digital assistant with the ability to process audio and visual information in real time. But the release of some of those features has been delayed as OpenAI works to address potential safety issues.
GPT-4o mini offers some of the same functionality and over time OpenAI plans to add more. AI companies, including Anthropic and Alphabet’s Google, frequently release nimbler, less capable and lower-cost versions of their most advanced models as a way to give developers more options. A smaller model, for instance, might be best for automating high-volume, basic tasks, while a larger model might field more complicated work. Some developers may want to use both in one application.
“In our mission to enable the bleeding edge, to build the most powerful, useful applications, we of course want to continue doing the frontier models, pushing the envelope here,” said Olivier Godement, the head of product for OpenAI’s API. “But we also want to have the best small models out there.”
Over the past week, some developers have been trying out the model, Godement said. E-mail start-up Superhuman used it to craft automated replies to messages, and financial services start-up Ramp relied on it to pull information from receipts.
New safety tactic
Initially, GPT-4o mini will be able to take in and generate text and images. Eventually, OpenAI said it will be able to process other types of content.
OpenAI also said that GPT-4o mini is the company’s first AI model to use a new safety tactic it developed called “instruction hierarchy”. The aim with this approach is to make an AI system prioritize some instructions — such as those from a company — over others in order to make it harder for people to get the tool to do things it shouldn’t. — Rachel Metz, (c) 2024 Bloomberg LP
