ChatGPT creator OpenAI debuted a cheaper and faster artificial intelligence (AI) model, GPT-4o mini, a smaller version of its flagship product on July 18.

In a press release, the company emphasized that its new product is aligned with its commitment to making intelligence as broadly accessible as possible. “Today, we’re announcing GPT-4o mini, our most cost-efficient small model. We expect GPT-4o mini will significantly expand the range of applications built with AI by making intelligence much more affordable.”

The latest release is a scaled-down model of GPT-4o, initially launched in May with the promise of enhanced multimodal capabilities that can cater to up to 50 languages. Now, GPT-4o mini offers cost-efficiency and increased speed that developers may use for simple, high-volume tasks.

Great things sometimes do come cheap

GPT 4o-mini is currently available to all ChatGPT users, including free, Plus, and Team subscribers, at 15 cents per million input tokens and 60 cents per million output tokens. To put that into perspective, the price is worth around 2500 pages in a standard book.

OpenAI says this cost is more than 60% cheaper than GPT 3.5 Turbo, the company’s previous smallest model.

So far, the new AI model has a context window of 128,000 tokens and knowledge up to October 2023. It supports text and vision modalities in the API, but its developers are improving the product to be available in video, image, and audio, too. GPT-4o mini will also be rolled out to ChatGPT Enterprise users a week after its first launch.

“For every corner of the world to be empowered by AI, we need to make the models much more affordable. I think GPT-4o mini is a really big step forward in that direction,” Olivier Godement, head of Product API at OpenAI, told TechCrunch in an interview.

Small but terrible

In addition to price, the GPT-4o mini also beats the GPT 3.5 Turbo and other small models in performance based on reasoning tasks, math and coding proficiency, and multimodal reasoning.

OpenAI claimed that its latest product achieved a score of 82.0% on MMLU, a scale used to measure textual intelligence in reasoning tasks that incorporate both text and vision. This surpassed 77.9% of Gemini 1.5 Flash and 73.8% of Claude 3 Haiku.

Likewise, GPT-4o mini reached the highest scores at 87.0% and 87.2% on MGSM and HumanEval, respectively, which measured the math capabilities and coding performance of the AI models.

An evaluation of the models’ multimodal reasoning using MMMU also revealed the same trend, wherein GPT-4o mini topped Gemini 1.5 Flash and Claude 3 Haiku with a score of 59.4%.

These percentages indicate the product’s advanced and faster capabilities despite its small size, making it useful for a variety of purposes. GPT 4o-mini can be used to call multiple APIs, pass a large volume of context to a model, and interact with customers using real-time text responses, among other things.

At the end of its press release, OpenAI said that safety features were built into its models from the beginning and reinforced at every step of the development process. The company vowed that harmful information was filtered out in pre-training of GPT-4o mini and that the model was designed to follow safety policies in post-training.