OpenAI Reduces Prices and Improves 'Lazy' GPT-4 That Initially Refused to Function

OpenAI continuously refines its models and pricing, with today being no exception. The company has introduced several new models while reducing API access costs—these changes are particularly significant for developers but also indicate potential shifts in consumer options.

The model that most users engage with is GPT-3.5 Turbo, typically accessed via ChatGPT. This model has become an industry benchmark; if your responses don't match the quality of ChatGPT’s, it raises questions about your output. GPT-3.5 Turbo is favored for its lower costs and quicker performance compared to GPT-4 across many tasks. As a result, paying users will be pleased to learn that input prices have decreased by 50% and output costs by 25%, now at $0.0005 per thousand tokens for input and $0.0015 per thousand tokens for output.

As developers experiment with these APIs for text-heavy applications—such as analyzing entire papers or books—token consumption can quickly increase. With open-source and self-managed models catching up to OpenAI's capabilities, the company is motivated to keep its customers engaged. This trend toward lower pricing also reflects the ongoing optimization of their models and enhanced infrastructure.

Additionally, GPT-3.5 Turbo has received a new version update, labeled 0125 (reflecting today's date). This update brings "various improvements"; however, there seem to be few noteworthy changes specified, which is surprising given that the previous version was 0613.

The GPT-4 Turbo now features a new preview model for API users, also labeled 0125, which boasts an intriguing enhancement: this model performs tasks like code generation more thoroughly than its predecessor. It aims to minimize instances of "laziness" where the model fails to complete tasks effectively. Perhaps it’s picked up some habits from the developers it learned from—could AI be quietly quitting as well?

Currently, this model is still in preview mode. However, GPT-4 Turbo with vision, or GPT-4 V, is set to launch for general availability in the upcoming months.

Moreover, there are several new and upgraded text embedding models geared toward technical users. OpenAI has also introduced an updated version of its free moderation API, which detects potentially harmful content. If you're interested in enhancing your moderation capabilities, check out version 007 for your experiments.