As the basis of ChatGPT, GPT-3.5 is probably the most influential major language model of all. Now it will have to give way to a successor, which OpenAI has now presented. GPT-4o mini is the smaller version of GPT-4o, which OpenAI released in May. Like its big brother, it is designed as a multimodal model, but at the moment it is limited in this regard. It can now process graphical inputs via APIs. The future will produce image, video and audio outputs.
Advertisement
A multimodal model with lots of performance for the money
GPT-4o Mini is trained on data extended to October 2023. Although the context window is about eight times larger than GPT-3.5 Turbo with 128k tokens, it is still smaller than Anthropic’s smallest model, Cloud 3 Haiku. However, it can generate 16,000 output tokens, which is far more than most comparable models. Even Cloud Sonnet, which competes in the next size category, only achieves half of that. In terms of output speed, GPT-4o Mini also leads the most popular LLM with 166 tokens per second.
As expected, in the benchmarks published by OpenAI it also tops smaller models like Google’s Haiku or Gemini Flash, but the gap is not particularly large. OpenAI could really score points with its price-performance ratio, at least for the moment. Both Anthropic and Google have based themselves on GPT-3.5 price expectations and are therefore currently significantly more expensive.
In general AI benchmarks, GPT-4o currently ranks at the top among “small” large language models. But the distances are not particularly large.
(Image: OpenAI)
GPT-4o Mini charges 15 cents per million input tokens and 60 cents per million output tokens. This is about 60 percent less than its predecessor. For comparison: for the larger GPT-4.o, OpenAI charges $5 per million inputs and $15 per million output tokens. That is more than thirty times more.
According to OpenAI, GTP-4o is the company’s first AI model to use a mini instruction hierarchy. This technique lets the model prioritize certain instructions over others. It is intended to make prompt injection attacks, jailbreaks, or system prompt extraction more difficult for users who bypass the built-in changes or instructions provided by the system prompt.
(ulw)
How to fix the Windows crash that caused the global computing crash
