In the early morning of January 26,OpenAIOn the official websiteChatGPTThe model has been greatly updated, with two new large and small text embedding models, a new GPT-4Turbo model (fixing lazy behavior), a free audit model, and a new GPT-3.5Turbo model.APISignificant price cuts were made.
OpenAI will also launch a new API key and visual management method to help developers observe API usage more simply and intuitively, and set more detailed usage permissions for API keys.
It is worth mentioning thatThe new embedding model can provide technical support for knowledge retrieval in ChatGPT, Assistants API, and many retrieval-enhanced generative development tools.
New text embedding model
"AIGC Open Community" first briefly introduces the embedding model: Embedding is a string of numbers that represents concepts in content such as natural language or code.Embeddings also make it easier for machine learning models and other algorithms to understand the relationships between content and perform tasks such as classification, content retrieval, search, and recommendations..
At the same time, embedding is a core component of the GPT series of models, which is used to convert input text (words or characters) into numerical vectors, such as word embedding, position embedding, and context embedding. These vectors can represent rich information of the input data and provide a deeper semantic understanding.
This time, OpenAI released the small text embedding model text-embedding-3-small and the large text embedding model text-embedding-3-large. The main features and performance are as follows.
1) Better performance:According to the performance tests released by OpenAI, the average score of the commonly used benchmark for multi-lingual retrieval (MIRACL) of small text embedding models increased from 31.4% to 44.0%; the average score of the commonly used benchmark for English tasks (MTEB) increased from 61.0% to 62.3%.
The average score of the large text embedding model on MIRACL increased from 31.4% to 54.9%, and on MTEB, the average score increased from 61.0% to 64.6%. The overall performance is stronger than that of small text.
2) Support shortening the embedding process to save costs:Developers often incur more costs when using larger embeddings (for example, storing them in vector memory for retrieval) than smaller embeddings, consuming more AI computing power, memory, and storage space.
To help developers save costs,OpenAI allows developers to shorten the embedding model (remove some numbers from the end of the sequence) by passing a dimension API parameter without causing the embedding model to lose its representational properties..
For example, on the MTEB benchmark,The large text embedding model can be shortened to 256 size, but the performance is still better than the unshortened small text embedding model of size 1536.
3) API price:Although the new text embedding model is very powerful, OpenAI has significantly reduced the price of its API.The API price of the small text embedding model has been reduced by 5 times compared to the previous model, and the price is 0.00002 USD per 1,000 tokens.The price of the Big Text Embedding Model is $0.00013 per 1,000 tokens.
New GPT-4Turbo preview model
Since OpenAI released the GPT-4Turbo model, more than 70% GPT-4API customers have switched to GPT-4Turbo. This is because GPT-4Turbo can provide greater context and performance.
Now,OpenAI releases a new GPT-4Turbo preview model - gpt-4-0125-preview.
Compared with before,This model is better at tasks like code generation, and fixes lazy behavior that was of great concern to users, as well as bugs affecting non-English UTF-8 generation..
For those who wish to automatically upgrade toup to dateDevelopers of the GPT-4Turbo preview model will always refer to OpenAIup to dateA preview of GPT-4Turbo from .
Free Review Model
To help developers reduce the output of illegal content in ChatGPT and improve security, OpenAI provides a free audit model API.
In addition, OpenAI will release aStrongestThe large audit model text-moderation-007 further improves the security of the model.
API price cuts
Next week, OpenAI will launch a new model in the GPT-3.5Turbo series - gpt-3.5-turbo-0125, and at the same time significantly reduce the price of the API.
The input price of the new model has been reduced by 50%, and the price is 0.0005 US dollars per 1000 tokens; the output price has been reduced by 25%, and the price is 0.0015 US dollars per 1000 tokens..
The model also makes various functional improvements, including improving the accuracy of the required format for responses and fixing an issue that caused text encoding errors in non-English language function calls.
A new visual API management approach
To help developers manage APIs more efficiently, OpenAI provides two new management methods.
1) Developers can now assign detailed permissions to API keys from the API Keys page.For example, a secret can be assigned read-only access to support an internal tracking dashboard, or restricted to access only certain endpoints.
2) After enabling the tracking function, the usage details and export functions can now display API key level metrics.Therefore, developers only need to set up a separate API key for each feature, team, product or project, and can easily view detailed usage at each feature, team, product or project level.
In the coming months, OpenAI will further enhance developers’ ability to use, observe, and control APIs, which is very important for large enterprises.