OpenAI launches GPT-4o Mini, outperforming GPT-3.5 Turbo

OpenAI recently launched a new model GPT-4o mini, replacing the original GPT-3.5.

This model is faster, cheaper and smarter, and can be used even by users of the free version.
OpenAI says GPT-4o mini outperforms GPT-3.5 Turbo on multiple tasks, including text processing, multi-modal reasoning and mathematical programming capabilities.

Key advantages

In the LMSYS rankings, GPT-4o mini performs well in reading performance, has a context window of 128,000 tokens, and supports up to 16,000 output tokens per request. GPT-4o mini performs well in multiple benchmark tests:

• Reasoning task performance: On reasoning tasks involving text and vision, GPT-4o mini scored 82.0% on MMLU (Massive Multi-task Language Understanding), outperforming Gemini Flash and Claude Haiku.
• Mathematics and programming abilities: GPT-4o mini scored 87.0% on MGSM (mathematical reasoning test) and 87.2% on HumanEval (programming test), both outperforming competing models.
• Multi-modal reasoning: On MMMU (Multi-modal Reasoning Evaluation), GPT-4o mini scored 59.4%, leading other models.

Price and availability

GPT-4o mini is priced very competitively at only 15 cents per million input tokens and 60 cents per million output tokens, which is more than 60% cheaper than GPT-3.5 Turbo. The model is already available in the Assistant API, Chat Completions API, and Batch API and is available to free, Plus, and Team users of ChatGPT, with enterprise users getting access next week.

safety

OpenAI emphasized the security of GPT-4o mini, which passed strict filtering and human feedback reinforcement learning (RLHF) technology from pre-training to post-training. This model applies the "instruction level method" in the API for the first time, which effectively improves the model's ability to resist illegal cracking, prompt injection, and system prompt extraction, ensuring the reliability of its answers.

Application scenarios

GPT-4o mini supports a wide range of tasks with its low cost and low latency, such as:

• Parallelize multiple model calls
• Provide lots of contextual information
• Quick, instant text responses to interact with customers