Blog

OpenAI is releasing a lighter, cheaper model for developers to tinker with, called the GPT-4o Mini. It costs significantly less than the full-size models and is said to be more capable than the GPT-3.5.

Building apps using OpenAI’s models can add up to a hefty bill. Developers who can’t afford to fiddle with it can ditch it altogether and opt for cheaper models like Google’s Gemini 1.5 Flash or Anthropic’s Claude 3 Haiku. Now OpenAI is entering the game with a lightweight model.

“I think the GPT-4o Mini really addresses OpenAI’s mission to make AI more widely accessible to people. If we want AI to benefit every corner of the world, every industry, every application, we need to make AI much more accessible,” said Olivier Godeman, who leads API Platform Product. On the edge.

Starting today, ChatGPT users on Free, Plus and Team plans can use GPT-4o Mini instead of GPT-3.5 Turbo, with Enterprise users gaining access next week. This means that GPT-3.5 will no longer be an option for ChatGPT users, but will still be available to developers via the API if they prefer not to switch to GPT-4o Mini. Godement has said that GPT-3.5 will be retired from the API at some point – they’re just not sure when.

The new, lightweight model will also support text and vision in the API, and the company says it will soon handle all multimodal inputs and outputs such as video and audio. With all these capabilities, it can seem like more capable virtual assistants that can understand your travel itinerary and create suggestions. However, the model is designed for simple tasks, so no one is exactly making Siri cheap.

This new model scored 82 percent on Measuring Massive Multitask Language Understanding (MMLU), a benchmark exam consisting of about 16,000 multiple-choice questions in 57 academic subjects. When the MMLU was first introduced in 2020, most mocks did quite poorly in it, which was the goal, as the mocks had become too advanced for previous benchmark exams. The GPT-3.5 scored 70 percent on this benchmark, the GPT-4o scored 88.7 percent, and Google claims the Gemini Ultra has the highest score of 90 percent. In comparison, competing models Claude 3 Haiku and Gemini 1.5 Flash scored 75.2% and 78.9% respectively.

It’s worth noting that researchers are wary of benchmarking tests like the MMLU because the way it’s administered varies slightly from company to company. This makes the results of different models difficult to compare, because New York Times reported. There’s also the problem that the AI ​​potentially has these answers in its data set, essentially allowing it to cheat, and there are usually no third-party raters involved in the process.

For developers looking to build low-cost AI applications, the launch of the GPT-4o Mini gives them another tool to add to their inventory. OpenAI allowed fintech startup Ramp to test the model using the GPT-4o Mini to create a tool that extracts cost data from receipts. So instead of digging through text fields, the user can upload a photo of their receipt and the model sorts everything for them. Superhuman, an email client, also tested the GPT-4o Mini and used it to create an auto-suggest feature for email replies.

The goal is to provide something lightweight and cheap for developers to build all the apps and tools they couldn’t afford to do with a larger, more expensive model like GPT-4. Many developers would turn to Claude 3 Haiku or Gemini 1.5 Flash before paying the impressive computing costs required to run one of the most robust models.

So what took OpenAI so long? Godement said this was “pure prioritization” as the company focused on making bigger and better models like GPT-4, which required a lot of “human and computer effort”. Over time, OpenAI noticed a trend of developers wanting to use smaller models, so the company decided now was the time to invest its resources in building the GPT-4o Mini.

“I think it’s going to be very popular,” Godement said. “Both from existing apps that use all of OpenAI’s AI, and from many apps that were released from the prices before.”