OpenAI, amidst intensifying competition in the generative AI field, has unveiled new advancements in its text-generating models, while simultaneously implementing price reductions.
The latest announcement from OpenAI introduces the launch of GPT-3.5-turbo and GPT-4, both of which incorporate cutting-edge text-generating capabilities. Notably, GPT-4 comes equipped with a feature known as function calling, enabling developers to provide programming functions to the AI models and have them generate corresponding code for execution.
This new function-calling capability has diverse applications, such as the creation of chatbots that utilize external tools to respond to inquiries, the conversion of natural language into database queries, and the extraction of structured data from text. OpenAI emphasizes that the models have been specifically fine-tuned to identify instances where function calling is necessary, responding with JSON that adheres to the function signature. This functionality offers developers a more reliable means of obtaining structured data from the models.
Expanding beyond function calling, OpenAI is introducing an enhanced version of GPT-3.5-turbo featuring a significantly increased context window. The context window represents the amount of text considered by the model before generating subsequent text and is measured in tokens. Models with smaller context windows tend to stray off-topic or forget the content of recent conversations, often leading to problematic outcomes.
The upgraded GPT-3.5-turbo now offers a context length four times larger than the original version, amounting to 16,000 tokens. However, this expanded capacity comes at a higher price, with rates set at $0.003 per 1,000 input tokens and $0.004 per 1,000 output tokens. OpenAI notes that this new version can process around 20 pages of text in a single operation, although it falls short of the capabilities demonstrated by Anthropic's flagship model, which can handle hundreds of pages. Additionally, OpenAI is currently conducting limited-release tests on a GPT-4 variant with a 32,000-token context window.
On the pricing front, OpenAI has implemented a 25% reduction for the original GPT-3.5-turbo version, without the expanded context window. Developers can now utilize this model for $0.0015 per 1,000 input tokens and $0.002 per 1,000 output tokens, equating to approximately 700 pages per dollar.
OpenAI has also reduced the pricing for text-embedding-ada-002, a popular text-embedding model. Text embeddings measure the semantic relatedness between different text strings and find applications in search engines and recommendation systems. The cost for text-embedding-ada-002 has been slashed by 75%, now priced at $0.0001 per 1,000 tokens. OpenAI attributes this reduction to enhanced efficiency in its systems, an area of significant focus for the company, given its substantial investments in research and infrastructure.
Following the release of GPT-4 in early March, OpenAI has indicated that its focus lies in providing incremental updates to existing models, as opposed to developing entirely new models from scratch. CEO Sam Altman recently reaffirmed this approach at a conference, emphasizing that OpenAI has not commenced training the successor to GPT-4, indicating that the company still has significant groundwork to cover before embarking on that endeavor.