Jan 26, 2024 11:50:00

OpenAI announces API update, improves GPT-4 task completion rate, adds new model, and features that allow you to track usage by API key, etc.

OpenAI, which develops and operates ChatGPT, has started providing a new model via API, and has also announced API-related updates such as updating the API usage management tool and reducing the price of GPT-3.5 Turbo.

New embedding models and API updates

https://openai.com/blog/new-embedding-models-and-api-updates

◆A new embedded model is now available
An 'embedding' is a series of numbers that represents a concept within the content, such as text or code. By quantifying information, algorithms such as machine learning models can understand the relationships between content, making it easier to perform tasks such as 'clustering' and searching, which groups related content together.

In this announcement, the embedding model, which previously only had one model, has been divided into small and large. ” model has appeared.

The small model has improved performance compared to the previous 'text-embedding-ada-002' model, such as increasing the average score of MIRACL, a benchmark commonly used in multilingual search, from 31.4% to 44.0%. While achieving this, we have also succeeded in reducing the price to 1/5th, or 0.00002 dollars (approximately 0.003 yen) per 1000 tokens.

On the other hand, the large model has increased the number of embedding dimensions to 3072, and is said to be a high-performance model with an average MIRACL score of 54.9%. The price is 0.00013 dollars (about 0.02 yen) per 1000 tokens, which is 1.3 times the previous price.

Additionally, both text-embedding-3 models now support the ability to reduce the number of embedding dimensions to reduce storage and memory usage. The scores for several dimensions of MTEB, a benchmark commonly used in English language tasks, are shown in the table below. It is possible to create embeddings of up to 1536 dimensions for small models and up to 3072 dimensions for large models.

	ada v2	text-embedding-3-small		text-embedding-3-large
Embed size	1536	512	1536	256	1024	3072
Average MTEB score	61.0	61.6	62.3	62.0	64.1	64.6

Please note that there are no plans to abolish the previous model text-embedding-ada-002, so although it says ``Use of the new model is recommended,'' the previous model can still be used.

◆GPT-3.5 Turbo model update
The GPT-3.5 Turbo model will be updated within a week and will be called 'gpt-3.5-turbo-0125'. At the same time, the price was reduced for the third time in the past year, with the input price being half the previous price of $0.0005 (approximately 0.074 yen) per 1000 tokens, and the output price being 25% cheaper at $0.0015 (approximately 0.22 yen) per 1000 tokens. ). In addition to lowering the price, it is said that various improvements have been made, such as improving response accuracy and fixing problems with text encoding.

◆Updated GPT-4 Turbo preview
A new preview model of GPT-4 Turbo 'gpt-4-0125-preview' was released at the same time as the announcement. Since around December 2023,

reports have been received that ``GPT-4 has stopped completing tasks'', but according to OpenAI, the task completion rate has improved with this update. The public release of GPT-4 Turbo is said to be 'within the next few months.'

◆Moderation model will be updated
OpenAI provides a free moderation API that quantifies the harmfulness of text. In this announcement, the most secure moderation model to date, 'text-moderation-007', was released. How to use the moderation API is described in the OpenAI documentation .

◆Updating management tools
You can now assign permissions to API keys, making them read-only or accessible only to specific endpoints. In addition, it is now possible to track usage status for each API key, and by simply preparing an individual API key, you can view usage status by any unit such as function, team, product, project, etc. Masu.

In addition, although individual usage tracking is automatically enabled for newly created API keys, it is necessary to enable individual tracking on the dashboard for conventional API keys. The company says it will continue to improve its API key management capabilities over the coming months, especially for larger organizations.