OpenAI drops costs and fixes ‘lazy’ GPT-4 that refused to work

OpenAI is all the time making slight changes to its fashions and pricing, and at the moment brings simply such an event. The corporate has released a handful of new models and dropped the value of API entry — that is primarily of curiosity to builders, but additionally serves as a bellwether for future client choices.

GPT-3.5 Turbo is the mannequin most individuals work together with, normally by means of ChatGPT, and it serves as a type of trade customary now — in case your solutions aren’t pretty much as good as ChatGPT’s, why hassle? It’s additionally a preferred API, being decrease price and quicker than GPT-4 on lots of duties. So paying customers shall 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 individuals play with utilizing these APIs for text-intensive purposes, like analyzing whole 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 verify 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. at the moment’s date) that features “various improvements” however apparently little that OpenAI thought value mentioning. The final model was 0613, so it’s slightly 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 technology extra totally than the earlier preview mannequin and is meant to scale back circumstances of “laziness” the place the mannequin doesn’t full a process.

Maybe the mannequin discovered the unsuitable classes from the builders whose habits it ingested and embodied! Possibly 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 common availability “in the coming months.”

There are additionally a handful of latest and improved textual content embedding fashions, that are extra for the individuals on the technical aspect. 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 the event you’re experimenting with utilizing this to reinforce your moderation wants.