OpenAI has unveiled its latest innovation, the GPT-4o Mini, marking its most budget-friendly small AI model to date. This new release aims to replace the GPT-3.5 Turbo as the default model for ChatGPT users, offering significant cost savings and improved efficiency. Although the company has not disclosed the parameter size, the GPT-4o Mini is comparable to other small language models (SLM) in the market. The model is now available to users as well as enterprises and developers through various application programming interfaces (API).
OpenAI Launches GPT-4o Mini
In a recent blog post, OpenAI highlighted the affordability of the GPT-4o Mini. This model is reported to be 60 percent less expensive than the GPT-3.5 Turbo, making it a highly economical choice for users. To put it in perspective, processing one million input tokens costs $0.15 (approximately Rs. 12), and one million output tokens cost $0.60 (around Rs. 50).
The GPT-4o Mini stands out for its low latency, offering quick response times, and currently supports text and vision processing in the API. OpenAI has future plans to extend its capabilities to include text, image, video, and audio for both input and output.
Despite not disclosing the parameter size, OpenAI shared that GPT-4o Mini features a context window of 128,000 tokens and can handle up to 16,000 output tokens per API request. The model’s knowledge cut-off is set to October 2023. In terms of performance, the GPT-4o Mini achieved an 82 percent score on the Massive Multitask Language Understanding (MMLU) benchmark, surpassing the earliest GPT-4 model on the Large Model Systems Organization’s (LMSYS) leaderboard, although these changes were not reflected on the website at the time of writing.
OpenAI’s internal testing suggests that GPT-4o Mini outperforms competitors like Gemini Flash and Claude Haiku in reasoning, text, and vision-based tasks. It has also scored 87 percent on the Multilingual Grade School Math (MGSM) benchmark and 87.2 percent on the HumanEval benchmark, demonstrating its math and coding proficiency, respectively.
For safety and reliability, OpenAI has implemented both automated and human evaluations following its Preparedness Framework, collaborating with 70 external experts across various disciplines to identify and mitigate potential risks.
The GPT-4o Mini is now the default model for all ChatGPT users, including free, Plus, and Team tiers, and will be available to enterprise users starting next week. The model is accessible as a text and vision API through the Assistants API, Chat Completions API, and Batch API, with fine-tuned versions expected in the future.
