OpenAI is committed to making intelligence accessible to all, and GPT-4o mini is a huge exemplification of this very mission. Such a small, affordable model will move mountains in AI applications since it brings advanced intelligence at a less expensive price and more widely available. The GPT-4o mini has achieved a remarkable score of 82% on the MMLU benchmark, outperforming GPT-4.1 in chat preferences as shown on the LMSYS leaderboard.
Unmatched affordability and performance
GPT-4o mini is priced at 15 cents per million input tokens and 60 cents per million output tokens. GPT-4o mini thus offers savings of more than 60% over the previous generation of models, including GPT-3.5 Turbo. On the other hand, this cost efficiency provides new opportunities to integrate AI into the operations of developers and businesses without having to break the bank.
Broad range of applications
The low-cost and low-latency GPT-4o Mini model is most fitting for the following use cases:
- Chaining or Parallelizing Multiple Model Calls: Enhances applications that require multiple API interactions.
- Handling Large Volumes of Context: Ideal for applications needing extensive context, such as codebases or conversation histories.
- Real-Time Customer Interactions: Powers fast, real-time text responses for customer support chatbots.
Versatile and forward-looking capabilities
Currently, GPT-4o mini supports text and vision in the API, while it will be expanded in the future to text, image, video, and audio inputs and outputs. A context window of 128K tokens; up to 16K output tokens per request for flawless performance on most tasks plus the improved tokenizer making it more efficient on non-English text.
Superior performance across benchmarks
GPT-4o mini performs well across several major benchmarks, including:
- Reasoning Tasks: Scores 82.0% on MMLU, outperforming Gemini Flash (77.9%) and Claude Haiku (73.8%).
- Math and Coding Proficiency: Achieves 87.0% on MGSM and 87.2% on HumanEval, surpassing other small models.
- Multimodal Reasoning: Scores 59.4% on MMMU, demonstrating superior performance in multimodal tasks.
Collaborative development and trusted partnerships
OpenAI tuned GPT-4o mini with the help of trusted partners such as Ramp and Superhuman. The companies witnessed the model significantly outperform GPT-3.5 Turbo in tasks ranging from the extraction of structured data off receipts to producing high-quality email responses.
Built-in safety measures
Safety at OpenAI is developed into the model development process. From pre-training to post-training, GPT-4o Mini has been truly infused with comprehensive measures to ensure its safety. The model filters undesirable content and addresses OpenAI’s policies using techniques such as reinforcement learning with human feedback.
GPT-4o mini has also undergone independent testing from third-party experts specializing in a variety of fields, including social psychology and misinformation, to further solidify its safety profile. It allows for the model to reject jailbreaks, prompt injections, or system prompt extractions using the instruction hierarchy method, lending to its reliability and confidence for large-scale applications.
Availability and pricing
GPT-4o mini is available through the Assistants API, Chat Completions API, and Batch API. This very powerful model is now available at a tenth of the cost; fine-tuning options will be added soon. Availability includes both Free and Plus ChatGPT users and Team and Enterprise users of GPT-4o mini. This will further OpenAI’s mission of democratizing AI.