OpenAI upgraded models for creating text amid growing competition in the field of generative artificial intelligence.
The company announced the release of new versions of GPT-3.5-turbo and GPT-4, the latter of which is the newest configuration of artificial intelligence that generates text with a feature called function calling.
A post published on the OpenAI blog contains information that the new feature allows developers to describe programming functions for GPT-3.5-turbo and GPT-4 and force models to create code to perform these functions. For example, it can create chatbots that answer questions using external tools, convert natural language into database queries, and extract structured information materials from text.
The new OpenAI solutions have been configured in such a way as to determine at what point there is a need to call a function. Also, the models are ready to respond using JSON, which corresponds to the signature of the function.
The new version of GPT-3.5-turbo received an expanded context window. This window, measured in tokens or raw text fragments, is considered by the model before generating any additional text material.
Models with small contextual windows often do not store data even about recent conversations, which is why there is a deviation from the topic.
GPT-3.5-turbo offers four times the length of the context (16,000 tokens) than a regular GPT-3.5-turbo, at twice the price — $0.003 for 1000 input tokens (injected into the model) and $0.004 for 1000 output tokens (generated by the model). OpenAI claims that this model can process about 20 pages of text at a time.
The company reduces prices for the original version of GPT-3.5-turbo without an extended context window by 25%. Developers will be able to use this model for $0.0015 for 1000 input tokens and $0.002 for 1000 output tokens, which is approximately 700 pages per dollar.
The cost of using text-embedding-ada-002, one of the most popular OpenAI text embedding models, is also reduced. Text attachments measure the relationship of strings and are commonly used for searches and recommendations.
Embedding text-ada-002 now costs $0.0001 per 1000 tokens, which is 75% lower than the previous price. OpenAI states that the downward revision of the price range has become possible as a result of increasing the efficiency of the company’s systems. The startup spends hundreds of millions of dollars on infrastructure, research, and development.
After the release of GPT-4, the company announced its intention to gradually update existing models, rather than create new ones. Sam Altman reported last week that OpenAI has not yet started training the successor to GPT-4, and noted that a lot of work should be done before launching this model.