OpenAI announce the launch of GPT-4o mini, the most budget-friendly small model to date. Aimed at making artificial intelligence more accessible, GPT-4o mini promises to revolutionize a variety of applications with its remarkable performance and affordability.
What is GPT-4o Mini?
GPT-4o mini is designed to deliver high intelligence at a low cost. It scores 82% on MMLU, outpacing GPT-4 in user preferences, as reported on the LMSYS leaderboard. Priced at just 15 cents per million input tokens and 60 cents per million output tokens, GPT-4o mini represents a significant cost reduction—over 60% cheaper than GPT-3.5 Turbo.
Target Applications
The model excels in various use cases such as:
- Chaining or parallelizing multiple model calls (e.g., making several API requests at once).
- Handling large volumes of context (e.g., accessing full code bases or conversation histories).
- Engaging in real-time, fast text responses for applications like customer support chatbots.
Features of GPT-4o Mini
Following are some of the features of GPT-4o Mini:
Multimodal Capabilities
Currently, GPT-4o mini supports both text and vision in its API, with future plans to include support for video and audio inputs and outputs. The model features an impressive 128K token context window and can generate up to 16K output tokens per request, ensuring it can handle expansive tasks effectively.
Enhanced Performance Metrics
GPT-4o mini has demonstrated superior capabilities across several benchmarks:
- Reasoning Tasks: Scoring 82.0% on MMLU, GPT-4o mini outperforms competitors like Gemini Flash (77.9%) and Claude Haiku (73.8%).
- Math and Coding Proficiency: With scores of 87.0% on MGSM (math reasoning) and 87.2% on HumanEval (coding tasks), GPT-4o mini exhibits exceptional skills compared to Gemini Flash (75.5% and 71.5%) and Claude Haiku (71.7% and 75.9%).
- Multimodal Reasoning: It scores 59.4% on MMMU, surpassing Gemini Flash at 56.1% and Claude Haiku at 50.2%.
Built-In Safety Measures
In the development of GPT-4o mini, OpenAI implement stringent practices to ensure the model's reliability.
- Pre-Training: Filter out harmful content such as hate speech, adult material, and spam during the pre-training phase.
- Post-Training Alignment: Utilize reinforcement learning from human feedback (RLHF) to fine-tune the model’s outputs, ensuring accuracy and adherence to OpenAI safety policies.
More than 70 external experts evaluated GPT-4o mini for potential risks, and OpenAI plan to share details in forthcoming documentation, including a system card and preparedness scorecard. The model also employs an instruction hierarchy method, enhancing its resistance against jailbreaks and prompt injections.
Availability and Pricing
GPT-4o mini is now accessible in various APIs, including the Assistants API, Chat Completions API, and Batch API. Developers can utilize the model at an incredibly low cost of 15 cents per million input tokens and 60 cents per million output tokens.
Starting today, ChatGPT Free, Plus, and Team users can leverage GPT-4o mini, replacing GPT-3.5. Enterprise users will gain access next week, aligning with OpenAI mission for widespread AI accessibility.
The Future of GPT-4o Mini and AI Accessibility
Since the introduction of text-davinci-003 in 2022, the cost per token for GPT-4o mini has decreased by an astonishing 99%, underscoring OpenAI commitment to driving down costs while enhancing the capabilities of the models.
Looking ahead, we anticipate a world where artificial intelligence seamlessly integrates into every application and website. GPT-4o mini is paving the way for developers to create powerful AI applications more efficiently and affordably.
“The future of AI is becoming more accessible, reliable, and embedded in our daily digital experiences,”.
For more detailed insights or to begin integrating GPT-4o mini into your applications, visit OpenAI official documentation.