GPT-4o Mini: A Revolution in Cost-Efficient AI Intelligence 🚀

Introduction

OpenAI has unveiled its latest innovation, GPT-4o mini, a small yet powerful model designed to make AI more accessible and affordable. This groundbreaking model is expected to expand the range of AI applications significantly, offering superior performance at a fraction of the cost of previous models.

Key Features of GPT-4o Mini

Cost-Efficiency and Accessibility

GPT-4o mini is priced at 15 cents per million input tokens and 60 cents per million output tokens, making it over 60% cheaper than GPT-3.5 Turbo. This affordability opens doors for various applications, including:

Chaining or parallelizing multiple model calls
Processing large volumes of context
Real-time text interactions for customer support

Enhanced Multimodal Capabilities

The model currently supports text and vision inputs through the API, with future plans to incorporate video and audio inputs. It has a context window of 128K tokens and can support up to 16K output tokens per request, making it versatile for extensive tasks.

Superior Performance Benchmarks

GPT-4o mini surpasses its predecessors and competitors in several benchmarks:

Reasoning Tasks: Scores 82.0% on MMLU, outperforming Gemini Flash and Claude Haiku.
Math and Coding Proficiency: Achieves 87.0% on MGSM and 87.2% on HumanEval, leading the pack in mathematical reasoning and coding tasks.
Multimodal Reasoning: Scores 59.4% on MMMU, showcasing its strong performance in multimodal evaluations.

Safety and Reliability

Safety is a cornerstone of GPT-4o mini’s design, featuring robust built-in measures such as filtering out undesirable content during pre-training and employing reinforcement learning with human feedback (RLHF) for post-training alignment. The model also uses an instruction hierarchy method to resist jailbreaks and prompt injections, enhancing the reliability and safety of responses.

Availability and Future Prospects

GPT-4o mini is now available in the Assistants API, Chat Completions API, and Batch API. It will be accessible to Free, Plus, and Team ChatGPT users starting today, with Enterprise access rolling out next week. Fine-tuning capabilities for GPT-4o mini will be introduced soon, further expanding its utility.

Conclusion

GPT-4o mini represents a significant step forward in making AI more accessible, reliable, and embedded in everyday digital experiences. OpenAI’s commitment to reducing costs while enhancing capabilities promises a future where powerful AI is integrated seamlessly into every app and website.

For more details and to start leveraging GPT-4o mini in your applications, visit OpenAI’s official website.

Author: OpenAI

Acknowledgments: Jacob Menick, Kevin Lu, Shengjia Zhao, Eric Wallace, Hongyu Ren, Haitang Hu, Nick Stathas, Felipe Petroski Such, Mianna Chen

Footnotes:

Evaluation numbers for GPT-4o mini are computed using the simple-evals repo with the API assistant system message prompt.

As of July 18th, 2024, an earlier version of GPT-4o mini outperforms GPT-4T 01-25.