GPT-4o Mini: Advancing Cost-Efficient Intelligence

OpenAI has introduced GPT-4o Mini, a cost-effective model aimed at providing advanced AI capabilities to a wider audience. This new model is priced significantly lower than its predecessors.

A Commitment to Accessible Intelligence

OpenAI aims to make AI broadly accessible. The launch of GPT-4o Mini supports this goal by offering an affordable yet powerful model. With a price of just 15 cents per million input tokens and 60 cents per million output tokens, GPT-4o Mini is much cheaper than previous models, offering substantial savings compared to GPT-3.5 Turbo.

Superior Performance at a Lower Cost

GPT-4o Mini delivers impressive performance despite its smaller size and lower cost. It achieves an 82% score on the MMLU (Massive Multi-Task Language Understanding) benchmark, outperforming GPT-3.5 Turbo on chat preferences according to the LMSYS leaderboard. Its low latency makes it ideal for applications requiring multiple model calls or real-time customer interactions.

Versatility and Multimodal Capabilities

GPT-4o Mini supports various input and output types, including text and vision. Currently, the API supports text and vision, with plans to add video and audio in the future. Its 128K token context window and capacity for 16K output tokens per request enable it to manage complex tasks, such as processing extensive code bases or lengthy conversation histories.

Benchmark Excellence

GPT-4o Mini stands out across multiple benchmarks, outperforming other small models like Gemini Flash and Claude Haiku in reasoning, math tasks, and coding performance. It scores 87.0% on the MGSM (math reasoning) and 87.2% on HumanEval (coding performance). In multimodal reasoning, it achieves 59.4% on the MMMU benchmark, demonstrating its strength in handling diverse data formats.

Enhanced Safety Measures

Safety is a critical aspect of AI deployment. GPT-4o Mini incorporates strict safety protocols similar to those used in GPT-4o, including content filtering during pre-training and behavior alignment through reinforcement learning with human feedback (RLHF). Over 70 external experts have evaluated GPT-4o to identify risks, ensuring that GPT-4o Mini is both safe and reliable for various applications.

Built for Developers

Developers will benefit from GPT-4o Mini's advanced function calling feature, which allows applications to efficiently fetch data or perform actions with external systems. This capability, paired with its long-context performance, makes GPT-4o Mini a robust tool for creating sophisticated AI-driven applications.

Real-World Applications

Trusted partners, including Ramp and Superhuman, have tested GPT-4o Mini. Their experiences show that it significantly outperforms GPT-3.5 Turbo in tasks like extracting data from receipt files and crafting quality email responses based on conversation history. This validation highlights the model’s practical advantages and its potential to foster innovation across various industries.

Availability and Pricing

GPT-4o Mini is accessible through several APIs, including the Assistants API, Chat Completions API, and Batch API. The pricing structure makes it affordable, with costs of 15 cents per million input tokens and 60 cents per million output tokens, equivalent to around 2500 pages in a standard book. Fine-tuning options will be available soon, allowing further customization for specific applications.

GPT-4o Mini brings significant improvements in cost-efficient intelligence. With its affordability and high performance, OpenAI is paving the way for a wider range of applications, enabling innovation across diverse sectors.

(Edited on September 4, 2024)