OpenAI releases GPT-4o-Mini, this one is much cheaper than GPT-4o.

GPT-4o costs $5/M input, $15/M output.
GPT-4o-mini costs just $0.15/M input, $0.60/M output.

That is, input is 33 times cheaper and output is 25 times cheaper.

But more importantly than the price is that the output context window has been increased to 16k output (128k input), which is superior to others on the market because the now famous Claude just upgraded its output context window from 4k to 8k last week. Which is still only half the size of the GPT-4o-mini.

This one's position is to replace GPT-3.5-turbo. Cheaper is better.

Most importantly, GPT-4o-mini is the first model to support heirarchy instructions, which is a way to prevent LLM from being jailbroken.

In common models on the market right now, the system prompt and user prompt have the same priority, allowing the user to instruct LLM to ignore commands in the system prompt, causing LLM to behave in a way that the service provider doesn't want it to, called jailbreak.

Having an instruction heirarchy, meaning different instruction (prompt) priority levels, allows LLM to better manage conflicts between system prompts and user prompts.

In terms of speed, testing with the API was approximately 150-200 tokens / sec, which is considered very fast.

As a result, LLM services will be safer, faster and cheaper.