OpenAI announced updates, including a more manipulable API model, function call capabilities, longer contexts, and lower prices.

OpenAI released gpt-3.5-turbo and gpt-4 earlier this year, and in just a few months, developers have built incredible applications based on these models.

Today, they're bringing some exciting updates:

▪ Added new function call capability in Chat Completions API.

▪ Newer and more controllable versions of gpt-4 and gpt-3.5-turbo.

▪ New 16k context version of gpt-3.5-turbo (vs. standard 4k version).

▪ The cost of OpenAI's advanced embedding models was reduced by 75%.

▪ The cost of input tokenization for gpt-3.5-turbo has been reduced by 25%.

▪ Announced the deprecation schedule for the gpt-3.5-turbo-0301 and gpt-4-0314 models.

All of these models offer the same data privacy and security guarantees that OpenAI introduced on March 1 — clients own all output generated from their requests, and their API data is not used for training.

function call

Developers can now describe functions to gpt-4-0613 and gpt-3.5-turbo-0613 and have the model intelligently choose to output a JSON object containing the arguments to call those functions. This is a new way to more reliably connect GPT's capabilities with external tools and APIs.

These models are fine-tuned to detect when a function needs to be called (depending on user input), and respond with JSON that conforms to the function's signature. Function calls enable developers to more reliably retrieve structured data from models.

For example, developers can:

▪ Create chatbots that answer questions by calling external tools (eg, similar to the ChatGPT plugin).

▪ Convert queries to function calls, such as "Email Anya to see if she wants to get coffee next Friday" to send_email(to: string, body: string), or "What's the weather like in Boston?" to get_current_weather(location: string, unit: 'celsius' | 'fahrenheit').

▪ Convert natural language into API calls or database queries, such as "Who are my top ten customers this month?" into an internal API call get_customers_by_revenue(start_date: string, end_date: string, limit: int), or "How many orders did Acme, Inc. place last month?" translates to an SQL query using sql_query(query: string).

▪ Extract structured data from text, eg define a function called extract_people_data(people: [{name: string, birthday: string, location: string}]) to extract all people mentioned in a Wikipedia article.

These use cases can be realized through new API parameters (functions and function_call) in the OpenAI /v1/chat/completions endpoint, developers can describe functions through JSON schema, and optionally ask the model to call a specific function. If a developer finds a situation where a function call could be improved, see the developer documentation and add evals to get started.

Since the Alpha release of the ChatGPT plugin, OpenAI has gained a lot of experience making tools and language models work together safely. However, there are still some unresolved research questions. For example, a proof-of-concept vulnerability example illustrates how untrusted data from tool output can instruct a model to perform unexpected actions. They are working to mitigate these and other risks. Developers can protect their applications by only using information from trusted tools and including user confirmation steps before performing actions with real-world impact, such as sending email, posting online, or purchasing.

new model

GPT-4

gpt-4-0613 includes an updated and improved model with function calls.

gpt-4-32k-0613 incorporates the same improvements as gpt-4-0613, while also adding a longer context length for better understanding of larger texts.

With these updates, OpenAI will be inviting more people on the waitlist to try GPT-4 in the coming weeks, with the intention of removing the waitlist entirely through this model. Thanks to everyone who has been patient, we look forward to seeing how you build applications with GPT-4!

GPT-3.5 Turbo

gpt-3.5-turbo-0613 includes the same function call capabilities as GPT-4, and more reliable maneuverability through system messages, both of which allow developers to more effectively guide the response of the model.

gpt-3.5-turbo-16k has 4x the context length of gpt-3.5-turbo and is twice as expensive: down to $0.003 per 1K input tokens and $0.004 per 1K output tokens. 16k contexts means the model can now support ~20 pages of text in a single request.

model deprecation

Today, OpenAI will begin the upgrade and deprecation process for the initial versions of gpt-4 and gpt-3.5-turbo that we announced in March. Applications using the stable model names (gpt-3.5-turbo, gpt-4, and gpt-4-32k) will be automatically upgraded to the new models above on June 27th. To compare model performance between different versions, our Evals library supports public and private evals to show the impact of model changes on your use case.

Developers who need more transition time can continue to use the old model by specifying gpt-3.5-turbo-0301, gpt-4-0314 or gpt-4-32k-0314 in the 'model' parameter of the API request. These old models will no longer be available after September 13, after which requests specifying these model names will fail. You can keep up to date with model deprecations on our model deprecations page. This is the first update to these models, so we're eagerly awaiting developer feedback to help us ensure a smooth transition.

lower pricing

OpenAI continues to improve the efficiency of the system and pass these savings back to developers, effective today.

embedded model

text-embedding-ada-002 is OpenAI's most popular embedding model. Today, their cost has been reduced by 75% to only $0.0001 per 1K markup.

GPT-3.5 Turbo

gpt-3.5-turbo is our most popular chat model, powering ChatGPT for millions of users. Today, OpenAI reduced the input labeling cost of gpt-3.5-turbo by 25%. Developers can now use the model for $0.0015 per 1K input tokens and $0.002 per 1K output tokens, which equates to approximately 700 pages of text per dollar.

gpt-3.5-turbo-16k is priced at $0.003 per 1K input token and $0.004 per 1K output token.