OpenAI AI models can now access plugins through the API. GPT-3.5 gets a larger context window.
OpenAI adds “functions” to the API for the gpt-4-0613 and gpt-3.5-turbo-0613 models. These function calls allow API-connected applications to more precisely steer GPT model capabilities, such as controlling external tools.
Both models have been enhanced to recognize when a function should be called based on user input, similar to ChatGPT with plugins enabled. An input like “what’s the weather like in the city” or “send an email to” is automatically converted into an appropriate function call like “send_email(to: string, body: string)”.
OpenAI points out existing security vulnerabilities when using plugins, where the output of an untrusted plugin can lead to unwanted actions by the language model. Developers can protect their applications by using only trusted tools and asking for user consent at certain points in the process, such as before sending an email.
GPT 3.5 gets larger context window
Another major update is the new gpt-3.5-turbo-061-16k model, which at 16k can handle four times more context, ie four times more sentence and word characters (tokens), than the traditional GPT-3.5 in ChatGPT. This allows users to reference more text at once in their queries, for example, to analyze a larger document at once, or to extract information from different parts of the document.
In addition, OpenAI is reducing the price of embeddings with the ada-002 model by 75 percent to $0.0001 per 1000 tokens and the price of GPT 3.5-turbo by 25 percent to $0.0015 per 1000 input tokens and $0.002 per 1000 output tokens. According to OpenAI, this is the equivalent of generating approximately 700 pages of text for one US dollar. The new 3.5 model with 16K costs $0.003 per 1K input tokens and $0.004 per 1K output tokens.
These price reductions may indicate that OpenAI will introduce more efficient and cheaper versions of GPT-4 in the near future, possibly GPT-4.5. There was also a price reduction for GPT-3 during the transition from GPT-3 to GPT-3.5.
Applications using the stable model names (gpt-3.5-turbo, gpt-4, and gpt-4-32k) will automatically be upgraded to the new models listed above on June 27th. You can check OpenAI’s Evals library to test if and how the model changes will impact your use cases.