Unleashing Cutting-Edge Artificial Intelligence on Budget with OpenAI’s GPT-4 Mini

OpenAI is launching a lightweight version of its flagship GPT-4o model. The new miniaturized model, aptly named “GPT-4o mini,” will be available for free and paid ChatGPT users later today.

Most AI companies, including Google and Meta, offer some kind of lightweight LLM. These tiny models are intended for simple or repetitive tasks. They don’t require a ton of processing power, they’re relatively inexpensive, and they’re pretty fast. So, they’re often the best choice for app or software integration.

GPT-4o is the most affordable and lightweight model in OpenAI’s arsenal. In fact, it’s “60% cheaper” than GPT-3.5 Turbo , and it outscores both Gemini 1.5 Flash and Claude 3 Haiku on the MMLU AI benchmark. Exact benchmark scores should be published on Artificial Analysis within the coming days.

For those wondering, the “60% cheaper” claim is totally accurate. Developers utilizing GPT-3.5 Turbo currently pay $0.50 for a million input tokens and $1.50 for a million output tokens. Transiting to GPT-4o mini brings per-million token pricing down to $0.15 and $0.60, respectively. That’s important because apps that currently offer AI functionality may improve, and developers that previously avoided AI due to cost concerns may be more inclined to join the bandwagon. The new model might also serve as the replacement for GPT-3.5 for ChatGPT users and developers, but that has not yet been confirmed.

Average ChatGPT users may see no reason to test GPT-4o mini, which is understandable, since it’s currently aimed at developers. However, free and paid ChatGPT users may still get some use out of GPT-4o mini, especially when generating images, proofreading several pages of text, or asking simple questions. A free user may choose the lightweight model to avoid hitting their daily GPT-4o usage cutoff, for example.

GPT-4o mini launches this afternoon (July 18th) and will be immediately available to all ChatGPT users. Developers also gain immediate access, though enterprise customers must wait an extra week.

Source: OpenAI via TechCrunch