
Designed to significantly expand the range of AI-powered applications, the GPT-4o mini costs just 15 cents per million input tokens and 60 cents per million output tokens – compared to the previous GPT-3.5 turbo model. More than 60% cheaper than
“Our mission is to make intelligence as widely accessible as possible,” said Sam Altman, CEO of OpenAI. “GPT-4o mini is an important step in this direction, offering developers and enterprises a cost-effective way to leverage the power of large language models.”
Despite its low price, the GPT-4o mini outperforms other small-scale models on the market, OpenAI said in a blog post. According to OpenAI’s internal benchmarks, the new model scored 82% on the Multimodal Language Understanding (MMLU) test, beating GeminiFlash (77.9%) and Claude Haiku (73.8%).
“The combination of cost efficiency and robust performance across multiple tasks makes the GPT-4o Mini a game changer,” said Dario Amudi, Head of Research at OpenAI. “We expect this model to enable a whole new wave of AI-powered applications and services.”
“The model’s ability to reason over text, vision and other modes opens up new use cases,” Altman said. “Developers can now build applications that seamlessly integrate natural language processing as well as AI-powered analysis of images, videos and audio.”
Improved safety and reliability
OpenAI has implemented various security measures in the GPT-4o mini, based on lessons learned from its deployment of the larger GPT-4o model.
The new model features improved filtering to remove potentially harmful content during pre-training and reinforcement learning techniques to align the model’s behavior with OpenAI’s security policies. Additionally, the GPT-4o mini is the first model to implement the company’s instruction classification method, which helps improve the model’s resistance to jailbreak and instant injection.
“We envision a future where AI-powered intelligence is seamlessly integrated into every app and website,” Altman said. “The GPT-4o mini is an important step towards this vision, making innovative, AI-powered solutions more accessible to developers.”
The model is now available through OpenAI’s Assistant API, Chat Completions API, and Batch API, with plans to roll out fine-tuning capabilities in the coming days. ChatGPT users will be able to access the GPT-4o mini starting today, replacing the previous GPT-3.5 turbo model.
Also read: Anthropic, Menlo Ventures Partners in $100 Million AI Fund