The new version of GPT-4o mini has reduced its price to $1 per million tokens, a move that puts OpenAI ahead of other AI companies in the competition.
OpenAI launched its most cost-effective model, GPT-4o mini, on the evening of July 18. The API input pricing for this model is 15 cents per million tokens, and the output price is 60 cents per million tokens, which is over 60% cheaper than GPT-3.5 Turbo. Despite the price reduction, GPT-4o mini's capabilities have significantly improved, achieving a good score of 82% in the MMLU test and surpassing GPT-4 in chat preferences on the LMSYS leaderboard.
Compared to small models from other companies, GPT-4o mini performs excellently in terms of cost-effectiveness:
- Google's Gemini 1.5 Flash: Input price $0.36 per million tokens, output price $1.10 per million tokens
- Anthropic's Claude 3 Haiku: Input price $0.26 per million tokens, output price $1.30 per million tokens
GPT-4o mini outperforms these competitors in all aspects of performance.
Domestic AI companies such as Deep Seek, Zhipu AI, ByteDance, Alibaba Cloud, Baidu, Tencent Cloud, and iFlytek also significantly reduced prices in June, but their price advantage has now been weakened. For example:
- Alibaba Cloud Qwen-Turbo: Input price $0.29 per million tokens, output price $0.87 per million tokens
- Baidu ERNIE 3.5 series: Input and output prices both $1.74 per million tokens
- ByteDance Doubao-pro-128k: Input price $0.72 per million tokens, output price $1.30 per million tokens
OpenAI CEO Sam Altman stated that GPT-4o mini is "moving towards intelligence so cheap it's unmeasurable." Compared to the GPT-3 text-davinci-003 model from 2022, the token cost of GPT-4o mini has decreased by 99%.
GPT-4o mini has the following features:
- 128k tokens context processing capability
- Supports text and visual input
- Supports 16k output tokens
- Will expand to video and audio input/output in the future
In multiple global authoritative benchmark tests, GPT-4o mini outperformed GPT-3.5 Turbo and other small models:
- MMLU text intelligence and reasoning test: 82.0%
- MGSM mathematical reasoning test: 87.0%
- HumanEval coding performance test: 87.2%
- MMMU multimodal reasoning evaluation: 59.4%
GPT-4o mini is now available for free on ChatGPT and is expected to gradually replace GPT-3.5. OpenAI plans to launch the fine-tuning feature for GPT-4o mini in the coming days.
AI expert Andrej Karpathy believes that the size competition for large language models is intensifying but showing a regressive trend. He predicts that very small but powerful models will emerge in the future.
OpenAI states that GPT-4o mini inherits the same strict safety protection mechanisms as GPT-4o and is the first model to apply OpenAI's instruction hierarchy approach, demonstrating stronger defense capabilities in the API.
The intensification of the price war for large models and the emergence of small, high-performance models will promote the wider integration of AI into daily scenarios while also prompting the industry to rethink how AI is implemented.