- November 7, 2023
- AI Projects
At its first developer conference, OpenAI introduced GPT-4 Turbo, an enhanced iteration of its primary text-generating AI model, GPT-4, touted as both more potent and cost-effective.
GPT-4 Turbo comes in two variants: one specializing in textual analysis and another proficient in comprehending both text and image context. The text-analyzing model is currently available in preview through an API, with OpenAI planning a general release “in the coming weeks.”
The pricing structure is set at $0.01 for every 1,000 input tokens (equivalent to about 750 words) and $0.03 for every 1,000 output tokens. Input tokens denote the bits of raw text fed into the model, while output tokens refer to the tokens generated based on the input tokens. As for the image-processing GPT-4 Turbo, the pricing will be contingent on the image’s dimensions. For instance, supplying a 1080×1080 pixel image to GPT-4 Turbo will cost $0.00765, according to OpenAI.
OpenAI declared, “We’ve enhanced the efficiency, enabling us to provide GPT-4 Turbo at a price that is three times lower for input tokens and twice as affordable for output tokens in comparison to GPT-4.”
GPT-4 Turbo features several enhancements over GPT-4, including a more recent knowledge base for more accurate responses to queries.
GPT-4 Turbo, like all language models, operates as a statistical tool for word prediction. It has been trained on a substantial number of examples from the web, enabling it to gauge word likelihood based on patterns and contextual nuances. For instance, if provided with a standard email ending in “Looking forward…,” GPT-4 Turbo can capably conclude it with “…to hearing back.”
While GPT-4 was trained up until September 2021, GPT-4 Turbo’s knowledge is up to April 2023, enhancing its capacity to provide precise answers about recent events occurring before the new cut-off date.
GPT-4 Turbo also introduces an expanded context window, measured in tokens, which determines the amount of text the model considers before generating additional text. With a substantial 128,000-token context window, GPT-4 Turbo outpaces even Anthropic’s Claude 2, making it the largest context window available in commercial models. This window equates to approximately 100,000 words or 300 pages, providing ample contextual depth for a variety of applications.
GPT-4 Turbo supports a new “JSON mode” that ensures the model’s responses conform to the valid JSON format, crucial for data transmission in web applications. Other parameters enable developers to elicit “consistent” completions more frequently and log probabilities for the most probable output tokens generated by GPT-4 Turbo, catering to niche applications.
OpenAI remarks, “GPT-4 Turbo performs better than our previous models on tasks that require careful adherence to instructions, such as generating specific formats (e.g., ‘always respond in XML’).” It is more adept at returning the correct function parameters.
OpenAI hasn’t neglected GPT-4 in its GPT-4 Turbo release. The company is launching an experimental access program for fine-tuning GPT-4, which will involve more oversight and guidance from OpenAI teams, given the technical complexities encountered. OpenAI acknowledges that preliminary results indicate that fine-tuning GPT-4 requires more effort to achieve meaningful improvements compared to GPT-3.5.
Lastly, OpenAI is doubling the tokens-per-minute rate limit for all paying GPT-4 customers, while retaining the pricing structure for different context window sizes.
