GPT-4 gets a new brother. GPT-4o Mini offers developers a lighter and cheaper alternative.

OpenAI introduces a new model of the GPT-4 family. Two months after the introduction of the GPT-4o, the GPT-4o Mini is now available. OpenAI describes this model as “lightweight”, which means that the model requires fewer computing resources to run. GPT-4o Mini is a direct competitor of Claude 3 Haiku and Google Gemini 1.5 Flash.

Because GPT-4o Mini uses less computing power, it provides developers with a cheaper platform to build AI applications. The price for using the GPT models is a complex calculation because it is based on the number of input and/or output tokens. For GPT-4o Mini, this is 15 cents per million input tokens and 60 cents per million output tokens, more than sixty percent cheaper than GPT-3.5.

Cheaper, but no less good

Cheaper and lighter doesn’t mean it’s less powerful. OpenAI has the necessary benchmarks to show that GPT-4o Mini is better than GPT-3.5, but also better than the competition, although AI benchmarks should be taken with a grain of salt. On the MMMU test, which assesses general knowledge, GPT-4o Mini scores 82 percent, and on the math and coding exams, the model scores 87 percent each.

GPT-4o Mini is available in ChatGPT today for both free and paid users, except for enterprise customers, whose turn will come next week. With the launch of GPT-4o Mini, GPT-3.5 will also be retired. The model has already disappeared from the ChatGPT selection window and OpenAI will also soon close API access to GPT-3.5.

Broadcom

The range of GPT models is becoming increasingly extensive. In order to increase capacity more quickly and to be more independent of Nvidia, OpenAI wants to build its own AI chips. Sam Altman is said to have approached Broadcom to agree on a collaboration on chip development, writes The Information based on anonymous sources.

OpenAI does not want to compete with Nvidia, but wants to be able to earn its own money. This has not been confirmed by either company for now.