OpenAI has revealed plans for GPT-4o mini, a new, cost-effective small model that provides high-quality results at a fraction of the price of its competitors. GPT-4o mini is set to be highly influential in a wide variety of fields, including customer support, software engineering, and creative writing, given its affordability and power.
The GPT-4o mini will charge only $0.15 and $0.60 per million input and output tokens respectively, a significant drop from the $5.00 and $15.00 charged for GPT-4o. Despite being a ‘mini’ version, the GPT-4o mini showcases notable capabilities, surpassing the GPT-3.5 Turbo on multiple benchmarks with the ability to handle both text and visual inputs.
GPT-4o mini has proven to outperform its competitors such as Google’s Gemini 1.5 Flash and Anthropic’s Claude 3 Haiku in the Massive Multitask Language Understanding (MMLU) benchmark, with a score of 82.0% compared to their scores of 77.9% and 73.8% respectively. Not only does it perform better, but GPT-4o mini is also intended to replace GPT-3.5 Turbo at no extra cost for ChatGPT Plus and Teams subscribers, making it even more attractive to potential users.
Startup companies like Ramp and Superhuman have already begun to see the benefits of using GPT-4o mini, citing successful results in tasks such as receipt categorization and tailored email responses. As interest in this affordable, efficient tool increases, OpenAI aims to expand the capabilities of GPT-4o mini, including the generation of imagery, audio, and video outputs. The model is also set to be available through Apple Intelligence this Fall with the release of iOS 18.
In the pursuit of ground-breaking innovation, OpenAI is also dedicated to ensuring the safety of its models. GPT-4o mini utilizes the same safety mechanisms developed for the larger GPT-4o. To further ensure the security and effectiveness of GPT-4o mini, the company invited over 70 experts specializing in fields such as social psychology and misinformation to test the model and help identify potential risks.
OpenAI has also implemented a new “instruction hierarchy” method, designed to enhance the model’s ability to resist jailbreaks, prompt injections, and system prompt extractions. This enhancement aims to increase the reliability of the model’s responses and reinforce its safety when being used on a large scale, an attractive feature to enterprise users who want to avoid any potential mistakes and illusions.
Despite the achievements of GPT-4o mini, the launch comes at a time when OpenAI has experienced some setbacks, such as the delay in their release of voice and emotion-reading features due to safety concerns. Nevertheless, OpenAI has demonstrated its commitment to continuous development and innovation amidst these challenges. With the release of GPT-4o mini, OpenAI continues to push the boundaries of pioneering technology, offering a cost-effective and high-performing model for a variety of applications.