OpenAI is all the time making slight changes to its fashions and pricing, and in the present day brings simply such an event. The corporate has launched a handful of latest fashions and dropped the worth of API entry — that is primarily of curiosity to builders, but in addition serves as a bellwether for future client choices.
GPT-3.5 Turbo is the mannequin most individuals work together with, often by way of ChatGPT, and it serves as a type of trade commonplace now — in case your solutions aren’t nearly as good as ChatGPT’s, why hassle? It’s additionally a well-liked API, being decrease price and sooner than GPT-4 on loads of duties. So paying customers might be happy to listen to that enter costs are dropping by 50% and output by 25%, to $0.0005 per thousand tokens in, and $0.0015 per thousand tokens out.
As folks play with utilizing these APIs for text-intensive functions, like analyzing total papers or books, these tokens actually begin to add up. And as open supply or self-managed fashions catch as much as OpenAI’s efficiency, the corporate wants to ensure its prospects don’t simply depart. Therefore the regular ratcheting down of costs — although it’s additionally a pure results of streamlining the fashions and bettering their infrastructure.
GPT-3.5 Turbo additionally will get a brand new mannequin model, 0125 (i.e. in the present day’s date) that features “various improvements” however apparently little that OpenAI thought price mentioning. The final model was 0613, so it’s a bit stunning they don’t point out extra.
GPT-4 Turbo will get a brand new preview mannequin for API use, additionally 0125, which has an attention-grabbing repair:
This mannequin completes duties like code era extra totally than the earlier preview mannequin and is meant to cut back circumstances of “laziness” the place the mannequin doesn’t full a job.
Maybe the mannequin discovered the improper classes from the builders whose habits it ingested and embodied! Perhaps AI is already quiet quitting. Good for AI!
That is nonetheless in preview mode. However GPT-4 Turbo with imaginative and prescient (that’s, GPT-4 V) will launch to basic availability “in the coming months.”
There are additionally a handful of latest and improved textual content embedding fashions, that are extra for the folks on the technical facet. And the corporate additionally launched a brand new model of its free moderation API — which identifies doubtlessly dangerous textual content. Search for model 007 in case you’re experimenting with utilizing this to reinforce your moderation wants.