GPT-4o mini - the new model from OpenAI and the small(er) Large Language Model revolution

OpenAI has just announced a new model called GPT-4o mini , which has the potential to greatly expand the future use of AI both directly via ChatGPT and in products built on OpenAI APIs. This development is part of a broader trend in AI towards cheaper and more efficient models. !ATTACH - {$originalname}.png

What it is

GPT-4o mini is a smaller, cheaper, and reasonably performant version of GPT-4o. Its key characteristics are:

• Very fast
• Very cheap
• Multimodal (it can interpret images with audio coming)
• Good enough for many tasks (performs well on many benchmarks)
• Much better than GPT 3.5 which it replaces

In my informal quick tests, it shows quite decent performance on many language tasks and even simple coding tasks. And it is blazingly fast. Other are reporting the same.

GPT-4o mini outperforms GPT-3.5 Turbo in textual intelligence (scoring 82% on MMLU compared to 69.8%) (OpenAI)

Availability

GPT-4o mini is now available to:

• Developers at a very low price (making development of some AI applications much cheaper)
• Rolling out to ChatGPT Free, Plus, and Teams
• Coming to Enterprise users in the coming weeks

GPT-4o mini is now available as a text and vision model in the Assistants API, Chat Completions API, and Batch API. Developers pay 15 cents per 1M input tokens and 60 cents per 1M output tokens (roughly the equivalent of 2500 pages in a standard book). We plan to roll out fine-tuning for GPT-4o mini in the coming days. (OpenAI)

Context

This release is part of a bigger trend of innovation in efficiency in the generative AI industry. We've now seen releases of cheap, fast, efficient models with acceptable performance from all 3 big AI Labs:

1. Claude 3 Haiku from Anthropic
2. Gemini 1.5 Pro Flash from Google
3. GPT-4o mini from OpenAI

This is also in the context of the release of many smaller open-source models that can run on more powerful computers locally with similar performance.

These models are not frontier models but are still far superior in Natural Language Processing to the best we had even 18 months ago. They all outperform GPT 3.5 and in select applications even match the performance of more advanced models.

What it means

Both this announcement and the broader trend in more efficient LLMs have the potential to broaden the impact of AI, especially in education:

• ChatGPT, Claude, and Gemini will be able to offer better free performance to more people
• We will see more AI products offer more for free within their freemium offering
• We may see more AI products in general as the barrier to entry to development with AI lowers even more
• The environmental impact of LLMs will continue to decrease as more tasks can be done with more energy-efficient models

Is GPT-4o better than GPT-4?

ChatGPT Plus users now see GPT-4 marked as legacy in the model selector. However, it's quite hard to measure LLM performance to that level of precision. Some reports and anecdotes suggest that GPT-4 is still the best model from OpenAI, but it is probably quite expensive to serve and energy-hungry. It's no surprise that OpenAI is pushing its successor, which has many other advantages such as full multimodality and also greater speed.