On January 25, OpenAI announced that it will introduce a new generation of embedded models, new GPT-4 Turbo and audit models, and new API usage management tools.。OpenAI also said it would soon reduce GPT-3..Price of 5 Turbo。

Two new embedding models

OpenAI says the updated ChatGPT will introduce two new embedding models: the small text embedding model text-embedding-3-small, and the large text embedding model text-embedding-3-large.。

These two new models have two characteristics compared to the previous generation (model text-embedding-ada-002 released in December 2022).。

One is better performance。OpenAI said that compared to the previous generation, the average score of text-embedding-3-small on its common benchmark for multilingual retrieval (MIRACL) increased from 31..4% to 44%, and the average score on the English Language Task (MTEB) common benchmark increased from 61% to 62.3%。

text-embedding-3-large is better than text-embedding-3-small in the above two performances。The average score of its MIRACL common benchmark is 54.9%, the average score of the MTEB common benchmark reached 64.6%。

Second, lower pricing.。While performance goes further, text-embedding-3-small / large pricing is lower。OpenAI says text-embedding-3-small is priced five times lower than its previous generation, at 0 per 1000 tokens.$00002, while the text-embedding-3-large price per 1000 tokens is 0.00013 USD。

And the reason why it can be priced lower is because OpenAI shortens embedding and reduces costs.。

In general, using larger embedded models (such as storing them in vector storage for retrieval) is usually more costly and consumes more computation, memory, and storage than using smaller embedded models。For the two new embedding models, text-embedding-3-small / large, OpenAI uses a technique for training that allows developers to weigh the performance and cost of using embedding。Specifically, developers can shorten the embedding by passing in API parameters without the embedding losing its conceptual representation characteristics。For example, on the MTEB benchmark, the text-embedding-3-large embedding can be shortened to 256 while still outperforming the unshortened embedding model text-embedding-ada-002 of size 1536。

This makes developers very flexible to use。For example, when using a vector datastore that only supports embeddings up to 1024 dimensions in length, developers can now still use text-embedding-3-large and specify a value of 1024 for the dimension API parameter, which will shorten the embedding from 3072 dimensions, reducing some precision in exchange for a smaller vector。

However, OpenAI also mentioned that the company does not intend to abandon the text-embedding-ada-002, they still welcome customers to continue to use the previous generation model。

Updated GPT-3.5 Turbo models

OpenAI says it will launch a new GPT-3 next week.5 Turbo model - gpt-3.5-turbo-0125。and will lower GPT-3 again.5 Turbo's price to help it expand customer size。The input price of the new model is 0 per 1000 tokens.$0005, 50% lower than before; output price is 0 per 1000 tokens.$0015, 25% lower than before。

OpenAI said the model will also make various improvements, including the format of responding to requests with higher precision, and fixing bugs that cause text encoding problems for non-English function calls。

New GPT-4 Turbo Preview Model

OpenAI also released a new GPT-4 Turbo preview model - gpt-4-0125-preview。This model completes tasks such as code generation better than the previous preview model, which can reduce the "laziness" of unfinished tasks of the model.。The new model also fixes bug issues affecting non-English UTF-8 generation。

For those who want to automatically upgrade to the new GPT-4 Turbo preview version, OpenAI also introduces a new model name alias, gpt-4-turbo-preview。The alias will always point to its latest GPT-4 Turbo preview model。

OpenAI says more than 70% of GPT-4 API customers have successfully converted to GPT-4 Turbo since the launch of GPT-4 Turbo。OpenAI next plans to fully launch GPT-4 Turbo, and hopes to complete the full market in the next few months。

In addition, OpenAI said it will launch its most powerful audit model to date, text-modernization-007, to advance its security efforts.。

API Update

To give developers a clearer picture of their API usage and control over API keys, OpenAI is rolling out two new initiatives。

One is that developers can now assign permissions to API keys from the API Keys page。For example, you can assign read-only access to keys to support internal tracking dashboards, or restrict access to only certain endpoints。

Second, after turning on the tracking function, the usage dashboard and usage export function can now display API key level indicators。This makes it easy to view usage at each feature, team, product, or project level by setting up a separate API key for each feature, team, product, or project level。

In the coming months, OpenAI plans to further improve developers' ability to view their API usage and manage API keys, especially in large organizations。