OpenAI Launches GPT-4o mini: High Performance at a Fraction of the Cost
GPT-4o mini achieved an impressive 82% score on the MMLU benchmark test, outperforming other small models in its class. The model features a 128,000-token context window and supports text and vision capabilities, with plans to add audio and video functionalities in the future.
The model's pricing is particularly cost-effective: $0.15 per million input tokens and $0.60 per million output tokens, making it over 60% cheaper than GPT-3.5 Turbo. Additionally, GPT-4o mini shows significant improvements in multilingual understanding, supporting numerous non-English languages.
GPT-4o mini Technical Specifications
GPT-4o mini's impressive technical specifications position it as a powerful yet cost-effective AI model. Here’s a comparison of key features between GPT-4o mini and other OpenAI models:
| Feature | GPT-4o mini | GPT-3.5 Turbo | GPT-4o |
|---|---|---|---|
| MMLU Score | 82% | 69.8% | 88.7% |
| Context Window | 128,000 tokens | 16,000 tokens | 128,000 tokens |
| Input Token Price | $0.15 / million | $0.50 / million | $5.00 / million |
| Output Token Price | $0.60 / million | $1.50 / million | $15.00 / million |
| Modalities | Text, Vision | Text | Text, Vision, Audio* |
| Knowledge Cutoff | October 2023 | September 2021 | October 2023 |
*GPT-4o's full multimodality includes audio, though API features may vary.
Capabilities and Applications
GPT-4o mini surpasses GPT-3.5 Turbo in text intelligence and multimodal reasoning while offering a substantially larger context window. It matches the context window size and knowledge cutoff of the flagship GPT-4o but at a fraction of the price. The model supports text and vision inputs, with future plans to incorporate audio and video capabilities, making it a versatile option for developers. Its enhanced multilingual understanding further broadens its utility across diverse applications and markets.
Accessibility and Integration
GPT-4o mini is immediately available in the OpenAI API suite, including the Assistants API, Chat Completions API, and Batch API. The model started rolling out to free and paid ChatGPT users (including Plus and Team subscribers) on July 18, 2024. Enterprise users were expected to gain access the following week. The model is also being integrated into the Microsoft Azure AI platform, allowing customers to leverage its capabilities for various applications, including audio, vision, and text processing.
Competitive Edge and Impacts
GPT-4o mini enters a competitive landscape as a strong contender against other small AI models. It outperforms Anthropic's Claude 3 Haiku on the multimodal reasoning benchmark (MMMU), scoring 59.4% compared to Haiku's 50.2%. GPT-4o mini also performs better than Google's Gemini Flash on the MMMU benchmark (59.4% vs. 56.1%). In terms of general intelligence, GPT-4o mini's 82% score on the MMLU benchmark is also notable, significantly surpassing GPT-3.5 Turbo's 69.8%. This performance, combined with its substantially lower price point and expanded context window, makes GPT-4o mini a highly competitive option for developers and businesses seeking cost-effective, high-performance AI solutions.
Future Prospects
GPT-4o mini is poised to make a significant impact on the AI field by making advanced language models more accessible and affordable. Its cost-effectiveness and improved performance are expected to drive wider adoption across various industries and applications. OpenAI envisions AI models seamlessly integrating into every application and website, and GPT-4o mini paves the way for developers to build and scale powerful AI applications more efficiently. The company remains committed to further reducing costs while enhancing model capabilities, having already achieved a 99% reduction in cost-per-token since the launch of text-davinci-003 in 2022. As GPT-4o mini becomes more widely adopted, it is likely to spur innovation in areas like customer service, content generation, and data analysis, potentially transforming how businesses and individuals interact with AI technology.