OpenAI launched GPT-4o mini, a smaller and cheaper LLM

OpenAI launched GPT-4o mini, a smaller and cheaper LLM
Artificial intelligence systems continue to evolve and OpenAI has introduced a new version of its LLM called GPT-4o mini. It is a smaller variant of the multimodal GPT-4o, while it is supposed to be a little smaller (but we don’t know exactly how much), but at the same time more affordable. It is supposed to be 60% cheaper against GPT-3.5 Turbo, which has powered some versions of the ChatGPT chatbot until today. Developers will pay 15 cents per million input tokens and 60 cents per million output tokens. In addition to the lower price, the advantage should be lower latency and, thanks to this, the easier possibility of parallel calls. Like GPT-4o, its mini version is also multimodal, so it can work with text, image, video and audio inputs.

It can handle up to 16k input tokens per command and the total popup can have up to 128k tokens. In terms of knowledge, the model is built on data up to October 2023. Although the GPT-4o mini is a lightweight version of the model, it outperforms most of the competition, including its own full-fledged GPT-3.5 Turbo. E.g. in reasoning in the MMLU benchmark it won 82,0 %, while Gemini Flash has 77.9% and Claude Haiku only 73.8%. The previous GPT-3.5 Turbo achieved only 69.8% in the same (however, the full-fledged GPT-4o has 88.7%).

V mathematics and writing source codes (MGSM benchmark) the novelty achieves excellent 87,0 % (GPT-4o even 90.5%), while Gemini Flash has only 75.5% and Claude Haiku 71.7%, the older GPT-3.5 Turbo was only at 56.3%. IN HumanEval it’s about 87,2 % for GPT-4o mini (90.2% for GPT-4o), Gemini Flash has 71.5%, Claude Haiku got to 75.9% and as usual, the original GPT-3.5 Turbo is already behind today with 68.0 %. GPT-4o mini also leads in multimodal thinking (MMMU) is scored 59,4 % (full version 69.1%). Competing Gemini Flash is slightly worse with 56.1%, Claude Haiku has 50.2%.

Also new to the model are new security features that better prevent hate speech, “adult” content, collection of personal data or spam, etc. OpenAI also worked on a new method to prevent attempts to jailbreak chatbots and various attempts to obtain system data of models.

Source: www.svethardware.cz