GPT-4 Turbo combines a longer context window with updated training data and lower stated API prices. OpenAI introduced two versions: one that analyzes text and another that can process both text and images. The announcement also included changes for developers using GPT-4.
More context and newer information
The text model’s knowledge cutoff is April 2023, compared with September 2021 for GPT-4. That gives GPT-4 Turbo a more recent basis for answering questions about events before its cutoff, though the cutoff still limits what it can know from training.
Its context window holds 128,000 tokens, four times the size of GPT-4’s, according to the article. A context window is the material a model can consider before it generates a response. A larger window can help it work with longer conversations or documents without losing track of earlier content.
The article equates 128,000 tokens with around 100,000 words or 300 pages. That capacity may be useful when a task depends on material spread across a lengthy text. It describes how much content can fit in context, rather than guaranteeing that every detail will be used correctly.
Lower prices and an image option
OpenAI said GPT-4 Turbo input tokens would cost $0.01 per 1,000 tokens and output tokens $0.03 per 1,000 tokens. Input tokens are supplied to the model; output tokens are generated in response. The company said the new rates were three times cheaper for input tokens and two times cheaper for output tokens compared with GPT-4.
The image-processing version has pricing based on image size. As one example, OpenAI said an image measuring 1080×1080 pixels would cost $0.00765 to pass to GPT-4 Turbo. The text-only version was available in API preview at announcement, while the company said it planned to make both versions generally available “in the coming weeks.”
Features aimed at developers
GPT-4 Turbo adds JSON mode, which is designed to ensure responses use valid JSON. That format is commonly used to pass structured data between parts of a web application. The announcement also described parameters intended to make completions more consistent and provide log probabilities for likely output tokens in specialized applications.
OpenAI said the model performs better than previous models at following instructions that specify a format, such as always responding in XML. It also said GPT-4 Turbo is more likely to return the right function parameters. These features target workflows where a model’s response must follow a predictable structure for software to use it.
GPT-4 fine-tuning and usage limits
Alongside the Turbo launch, OpenAI opened an experimental access program for fine-tuning GPT-4. Fine-tuning adapts a base model for particular needs. The company said the program would involve more oversight and guidance from its teams than the GPT-3.5 fine-tuning program, citing technical hurdles.
OpenAI’s preliminary results indicated that GPT-4 fine-tuning required more work to produce meaningful improvements over the base model than GPT-3.5 fine-tuning, which had delivered substantial gains. The experimental program therefore came with a more hands-on approach as the company explored what fine-tuning could achieve.
OpenAI also doubled the tokens-per-minute rate limit for all paying GPT-4 customers. The announcement said GPT-4 pricing would remain the same: $0.03 per input token and $0.06 per output token for the model with an 8,000-token context window, or $0.06 per input token and $0.012 per output token for GPT-4 with a 32,000-token context window.
Taken together, the changes addressed several developer concerns at once: how much material a model can consider, how recent its knowledge is, the cost of using it, and how reliably its output fits an application. GPT-4 Turbo was presented as the lower-cost option with expanded capabilities, while fine-tuning remained an experimental effort requiring further work.