OpenAI Launches GPT-4o Mini as Company’s Lowest Cost Small AI Model
OpenAI on Thursday released its latest artificial intelligence (AI) model, GPT-4o Mini. It’s a small AI model, and the company says it’s its most cost-effective model yet. While ChatGPT’s creator didn’t specify the model’s parameter size, it did stress that it’s comparable to other small language models (SLMs). GPT-4o Mini will replace GPT-3.5 Turbo on ChatGPT for all users as the default AI model. It will also be made available as an application programming interface (API) for businesses and developers.
OpenAI Releases GPT-4o Mini
In a blog postthe company announced the public release of its latest AI model. One of the key highlights of the new AI model is its cost-effectiveness. The company says it is 60 percent cheaper than GPT-3.5 Turbo, which was the previous cheapest small model. In real terms, processing one million input tokens costs the company $0.15 (roughly Rs. 12), and per million output tokens it costs $0.60 (roughly Rs. 50).
Because GPT-4o Mini is a small model, it also offers low latency (time taken to show the response), making it one of the company’s most efficient AI models. It currently supports text and vision (image processing) in the API, and OpenAI says that in the future it will support text, image, video, and audio as both input and output.
While the AI firm did not share the parameter size of the new SLM, it revealed that GPT-4o Mini has a context window of 1,28,000 tokens and can support up to 16,000 output tokens per request (on API). The knowledge limit will last until October 2023. Speaking of performance, OpenAI said that the AI model scored 82 percent on the Massive Multitask Language Understanding (MMLU) benchmark, outperforming GPT-4’s earliest model on independent AI chatbot ranking website Large Model Systems Organization’s (LMSYS) leaderboard. However, the changes have not been reflected in the website at the time of writing this.
OpenAI also claimed that the GPT-4o Mini outperformed Gemini Flash and Claude Haiku on reasoning-based text and vision-based tasks, based on internal testing. It also claims to score 87 percent on the Multilingual Grade School Math (MGSM) benchmark and 87.2 percent on the HumanEval benchmark. These two benchmarks measure the AI model’s math and coding skills, respectively.
For safety, the AI firm said it used both automated and human evaluations under its Preparedness Framework. The company also tested the AI model with 70 external experts from various disciplines to identify potential risks.
The GPT-4o Mini has been rolled out as the default AI model for ChatGPT for free, Plus, and Team users. The AI model will be extended to enterprise users starting next week. It is also available as a text and vision model in the Assistants API, Chat Completions API, and Batch API. Refined versions of the model will be released in the future.
