OpenAI Launches GPT-4o Mini API with 60% Price Drop, Making Smaller Models More Accessible

OpenAI Launches Cost-Effective GPT-4o Mini Model for AI Applications

On July 19, OpenAI officially introduced the GPT-4o mini model, touted as the most cost-effective small model available. This model is designed to replace GPT-3.5 Turbo, competing with Claude 3 Haiku and Gemini 1.5 Flash, and is expected to significantly reduce the costs associated with AI applications.

Achieving an 82% score on the MMLU benchmark and outperforming GPT-4 in the LMSYS chat scores, GPT-4o mini offers commercial pricing of $0.15 per million input tokens and $0.60 per million output tokens—more than 60% cheaper than GPT-3.5 Turbo.

The model supports text and visual inputs via API, with plans to expand to text, images, videos, and audio in the future. With a context window of 128K tokens and knowledge updated through October 2023, GPT-4o mini also benefits from an enhanced tokenizer shared with GPT-4o, making it more efficient in handling non-English text.

In academic benchmarks for text intelligence and multimodal reasoning, GPT-4o mini outperforms both GPT-3.5 Turbo and other small models, supporting the same range of languages as GPT-4o. Its advanced function calling capabilities simplify the development of applications that retrieve data or interact with external systems, improving the handling of long-context queries.

GPT-4o mini excels in several key benchmarks:

- Reasoning Tasks: Scoring 82.0%, surpassing Gemini Flash (77.9%) and Claude Haiku (73.8%).

- Mathematics and Coding Skills: Achieving 87.0% in the MGSM math reasoning test and 87.2% in the HumanEval coding performance, both exceeding Gemini Flash and Claude Haiku.

- Multimodal Reasoning: Scoring 59.4% in the MMMU assessment, outpacing competitors.

In terms of safety, GPT-4o mini integrates the same safety measures as GPT-4o. OpenAI conducted rigorous evaluations through automated and manual assessments, partnering with over 70 external experts to identify and mitigate potential risks, enhancing the model's safety profile.

As of today, GPT-4o mini is accessible through the Assistant API, Chat Completions API, and Batch API, with developers able to obtain access for a suitable fee. Starting now, GPT-4o mini is available to free, Plus, and Team users of ChatGPT, with enterprise users set to gain access next week.