GPT-4o mini from OpenAI’s small model is the AI company’s latest release and is expected to be a cheaper model alternative. The newly released model is now available for developers as a text and vision model. Enterprise users will be allowed access from next week onwards.
OpenAI has stated that their new model has scored 82% on MMLU. The model outdoes GPT-41 in terms of performance on chat preferences in the LMSYS leaderboard. The new model is marked at 15 cents per million input tokens and 60 cents per million output tokens which is significantly cheaper than other models.
GPT-4o and its features
The AI company states that the mini-model allows for a wider scope of tasks due to its latency and lower costs. Support for vision and text in the API is currently available, with plans to extend the support to image, text, audio and video outputs and inputs.
Furthermore, OpenAI’s new model has also been assessed through various benchmarks. As mentioned earlier, the mini model scores 82% on MMLU which is a step ahead of Gemini Flash’s 77.9% and Claude Haiku’s 73.8%.
Additionally, GPT-4o scores an impressive 87.0% on the MGSM scale which measures math reasoning. It scores 87.2% for coding proficiency on HumanEval. It also scores a 59.4% on MMMU, showing an impressive performance for multimodal reasoning.
Safety measures are implemented to prevent the model from learning inappropriate output. It also resists jailbreaks, system prompt extractions and prompt injections. RLHF (reinforcement learning with human feedback) has been implemented to improve the model’s accuracy of responses.
Suggested:
OpenAI GPT-4 Turbo is Paving the Way for a New Era in AI.
From OpenAI, ChatGPT Enterprise is Now Accessible.
Proton Scribe: Private Writing Assistant by Proton