OpenAI has revealed substantial upgrades to its GPT language models, including GPT-4 and GPT-3.5. These improvements encompass more extensive knowledge bases, significantly expanded context windows, and a commitment to shield customers against copyright litigation, taking a cue from Google and Microsoft.
A major unveiling from OpenAI is the GPT-4 Turbo, currently available through an API preview. Trained with data up to April 2023, this marks a significant leap from the previous GPT-4 version released in March, which was based on data only up to September 2021. The company plans to introduce a finalized Turbo model soon, although the specific date remains undisclosed.
GPT-4 Turbo offers a broad 128K context window, enhancing its understanding of inquiries and providing more nuanced responses compared to its predecessors with smaller context windows of 8K and 32K. OpenAI emphasizes the cost-effectiveness of GPT-4 Turbo for developers, with a lowered input cost of $0.01 per 1,000 tokens, down from $0.03 in previous versions. The new model is estimated to be three times more cost-effective.
This upgraded iteration preserves its capability to handle image prompts, support text-to-speech requests, and integrate with DALL-E 3, an addition announced earlier. The enhancements empower users to direct the model to perform more intricate tasks within a single prompt, including the option to specify the desired coding language, like XML or JSON.
GPT-3.5 Turbo, optimized for non-chat functions, also receives improvements in its context window, capabilities, and pricing. Offering a default 16K context window, the model shares functional updates with GPT-4 Turbo, priced at $0.01 for input and $0.002 for output.
OpenAI’s latest strides extend beyond model enhancements. They have introduced a Copyright Shield to defend customers facing legal copyright claims, covering features of ChatGPT Enterprise and the developer platform. OpenAI will intervene and cover the expenses incurred if customers encounter legal challenges concerning copyright issues.
