The AI vendor launched a small language model that is priced at 15 cents per million input tokens. It also introduced new compliance tools for ChatGPT Enterprise users.
OpenAI is following market trends with a new small language model as well as new compliance and security tools for ChatGPT Enterprise.
On Thursday, the AI vendor introduced GPT-4o mini and new controls in ChatGPT Enterprise for workspace admins. The GPT-4o mini is cheaper and faster than current OpenAI models, according to the AI vendor. It is priced at 15 cents per million input tokens and 60 cents per million output tokens.
GPT-4o mini currently supports text and vision. Support for text, image, video and audio inputs and outputs will be available in the future.
The model also has a context window of 128K tokens and supports up to 16K output tokens per request, OpenAI said. A context window is how much text, image or video a language model can process.
This expansion is a stark contrast to summer 2023, when vendors were focused on creating bigger and larger language models.
"This is all about exploiting the demand curve," said David Nicholson, an analyst at The Futurum Group. The current demand is shaped by what enterprises need, he added.
Previously, the focus was on large language models. However, enterprises are realizing they don't necessarily need the biggest language model to perform tasks due to the costs and power consumption levels that accompany larger language models.
The introduction of GPT-4o mini also confirms the trend that one model will not be the ruler of them all, said Mark Beccue, an analyst at TechTarget's Enterprise Strategy Group.
"Super-large models are for certain specific things, and smaller models are for other things," Beccue said.
He added that models are not only getting smaller, but they're also becoming more domain-specific and being used for specific use cases.
These domain-specific models can be assigned to specific tasks, and the smaller the model, the better they are for fine-tuning, Beccue added.
However, OpenAI's new model comes after the release of other small language models such as Microsoft Phi and Google Gemini Nano. It also follows other small open source models, such as Mistral Small.
It might seem like OpenAI is late to the market with this release, but that's not the case, Nicholson said.
"This is going to be sort of a leapfrog thing that's going to happen," he said. "The mini version of OpenAI is going to be more powerful than the full-blown version of previous generations of other people's LLMs and so forth."
GPT-4o mini is currently available to Plus and Team ChatGPT Free users. Enterprise users will have access starting next week.
New compliance and security tools
ChatGPT Enterprise users also have access to new compliance and administrative tools that OpenAI introduced on Thursday.
The new tools come one year after the release of ChatGPT Enterprise, which differs from the consumer version of ChatGPT because of its enterprise-grade security and controls.
The new administrative tools are meant to support enterprise customers with managing compliance and data security as well as scaling user access.
For example, Enterprise Compliance API lets admins audit their ChatGPT Enterprise workspace data. It provides time-stamped interactions of conversations, uploaded files, workspace GPT configuration and workspace users.
It also helps those in regulated industries, such as finance and healthcare, comply with logging and audit requirements.
Enterprise workspace owners can access the Enterprise Compliance API directly or use a third-party integration.
Workspace admins also have expanded GPT controls that let admins create an approved list of specific domains for granular controls over actions.
With the expanded controls, admins can create and edit user groups to control access and permissions. They can also manage GPT sharing permissions and approve specific third-party GPTs in their workspace.
"These kinds of settings are table stakes for an enterprise application," Beccue said. "They're late to the game."
He added that Microsoft Copilot has had these security guardrails on Copilot for the past year. For many enterprises, this should have been added with the introduction of ChatGPT Enterprise, Beccue said.
"These kinds of things are very important when you're doing enterprise-grade solutions," he said.
Moreover, as an organization that built its mission on AI safety, it's inconsistent for OpenAI to consider safety guardrails after introducing its product and not before, Beccue continued.
"If that's first and foremost what you do as a company, why are you doing it a year after you've introduced the product?" he added. "There's some conflict right in their mission -- that's clear."
This is all about exploiting the demand curve.
David NicholsonAnalyst, Futurum Group
However, the infancy of generative AI means one can't be critical of when vendors like OpenAI roll out security measures, Nicholson said.
"It's a moving target," he said. Having the security tools on Day 1 could mean they are easily hacked, he continued. "Because there is this huge potential, there is going to be more profit in breaking these systems [and] leveraging them with malice, with malintent."
One of the security tools that stands out is the Expanded GPT controls, which lets admins keep people out and audit access, Nicholson added.
"AI security is going to become a practice in and of itself," he said. "We are underestimating the damage that someone can do if they have surreptitious control over a generative AI in particular."
OpenAI will also roll out the System for Cross-domain Identity Management next week.
SCIM will let admins sync their internal employee directories with their ChatGPT Enterprise workspace.
Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.
