Skip to main content

January 25, 2024

New embedding models and API updates

We are launching a new generation of embedding models, new GPT-4 Turbo and moderation models, new API usage management tools, and soon, lower pricing on GPT-3.5 Turbo.

New Embeddings Models And API Updates

We are releasing new models, reducing prices for GPT-3.5 Turbo, and introducing new ways for developers to manage API keys and understand API usage. The new models include:

  • Two new embedding models

  • An updated GPT-4 Turbo preview model 

  • An updated GPT-3.5 Turbo model

  • An updated text moderation model

By default, data sent to the OpenAI API will not be used to train or improve OpenAI models.

New embedding models with lower pricing

We are introducing two new embedding models: a smaller and highly efficient text-embedding-3-small model, and a larger and more powerful text-embedding-3-large model.

An embedding is a sequence of numbers that represents the concepts within content such as natural language or code. Embeddings make it easy for machine learning models and other algorithms to understand the relationships between content and to perform tasks like clustering or retrieval. They power applications like knowledge retrieval in both ChatGPT and the Assistants API, and many retrieval augmented generation (RAG) developer tools.

Loading...

A new small text embedding model

text-embedding-3-small is our new highly efficient embedding model and provides a significant upgrade over its predecessor, the text-embedding-ada-002 model released in December 2022

Stronger performance. Comparing text-embedding-ada-002 to text-embedding-3-small, the average score on a commonly used benchmark for multi-language retrieval (MIRACL(opens in a new window)) has increased from 31.4% to 44.0%, while the average score on a commonly used benchmark for English tasks (MTEB(opens in a new window)) has increased from 61.0% to 62.3%.

Reduced price. text-embedding-3-small is also substantially more efficient than our previous generation text-embedding-ada-002 model. Pricing for text-embedding-3-small has therefore been reduced by 5X compared to text-embedding-ada-002, from a price per 1k tokens of $0.0001 to $0.00002.

We are not deprecating text-embedding-ada-002, so while we recommend the newer model, customers are welcome to continue using the previous generation model.

A new large text embedding model: text-embedding-3-large

text-embedding-3-large is our new next generation larger embedding model and creates embeddings with up to 3072 dimensions.

Stronger performance. text-embedding-3-large is our new best performing model. Comparing text-embedding-ada-002 to text-embedding-3-large: on MIRACL, the average score has increased from 31.4% to 54.9%, while on MTEB, the average score has increased from 61.0% to 64.6%.

Eval benchmark

ada v2

text-embedding-3-small

text-embedding-3-large

MIRACL average

31.4

44.0

54.9

MTEB average

61.0

62.3

64.6

text-embedding-3-large will be priced at $0.00013 / 1k tokens.

You can learn more about using the new embedding models in our Embeddings guide(opens in a new window).

Native support for shortening embeddings

Using larger embeddings, for example storing them in a vector store for retrieval, generally costs more and consumes more compute, memory and storage than using smaller embeddings.

Both of our new embedding models were trained with a techniqueA that allows developers to trade-off performance and cost of using embeddings. Specifically, developers can shorten embeddings (i.e. remove some numbers from the end of the sequence) without the embedding losing its concept-representing properties by passing in the dimensions API parameter. For example, on the MTEB benchmark, a text-embedding-3-large embedding can be shortened to a size of 256 while still outperforming an unshortened text-embedding-ada-002 embedding with a size of 1536.

Loading...

This enables very flexible usage. For example, when using a vector data store that only supports embeddings up to 1024 dimensions long, developers can now still use our best embedding model text-embedding-3-large and specify a value of 1024 for the dimensions API parameter, which will shorten the embedding down from 3072 dimensions, trading off some accuracy in exchange for the smaller vector size.

Other new models and lower pricing

Updated GPT-3.5 Turbo model and lower pricing

Next week we are introducing a new GPT-3.5 Turbo model, gpt-3.5-turbo-0125, and for the third time in the past year, we will be decreasing prices on GPT-3.5 Turbo to help our customers scale. Input prices for the new model are reduced by 50% to $0.0005 /1K tokens and output prices are reduced by 25% to $0.0015 /1K tokens. This model will also have various improvements including higher accuracy at responding in requested formats and a fix for a bug(opens in a new window) which caused a text encoding issue for non-English language function calls.

Customers using the unpinned gpt-3.5-turbo model alias will be automatically upgraded from gpt-3.5-turbo-0613 to gpt-3.5-turbo-0125 two weeks after this model launches.

Updated GPT-4 Turbo preview

Over 70% of requests from GPT-4 API customers have transitioned to GPT-4 Turbo since its release, as developers take advantage of its updated knowledge cutoff, larger 128k context windows, and lower prices. 

Today, we are releasing an updated GPT-4 Turbo preview model, gpt-4-0125-preview. This model completes tasks like code generation more thoroughly than the previous preview model and is intended to reduce cases of “laziness” where the model doesn’t complete a task. The new model also includes the fix for the bug impacting non-English UTF-8 generations.

For those who want to be automatically upgraded to new GPT-4 Turbo preview versions, we are also introducing a new gpt-4-turbo-preview model name alias, which will always point to our latest GPT-4 Turbo preview model. 

We plan to launch GPT-4 Turbo with vision in general availability in the coming months.

Updated moderation model

The free Moderation API allows developers to identify potentially harmful text. As part of our ongoing safety work, we are releasing text-moderation-007, our most robust moderation model to-date. The text-moderation-latest and text-moderation-stable aliases have been updated to point to it. You can learn more about building safe AI systems through our safety best practices guide(opens in a new window).

New ways to understand API usage and manage API keys

We are launching two platform improvements to give developers both more visibility into their usage and control over API keys.

First, developers can now assign permissions to API keys from the API keys page(opens in a new window). For example, a key could be assigned read-only access to power an internal tracking dashboard, or restricted to only access certain endpoints.

Second, the usage dashboard and usage export function now expose metrics on an API key level after turning on tracking(opens in a new window). This makes it simple to view usage on a per feature, team, product, or project level, simply by having separate API keys for each.

Image2

In the coming months, we plan to further improve the ability for developers to view their API usage and manage API keys, especially in larger organizations.

For the latest updates on OpenAI's APIs, follow us on X at @OpenAIDevs(opens in a new window).

Footnotes

Author

OpenAI

Acknowledgments

Juntang Zhuang, Paul Baltescu, Joy Jiao, Arvind Neelakantan, Andrew Braunstein, Jeff Harris, Logan Kilpatrick, Leher Pathak, Enoch Cheung, Ted Sanders, Yutian Liu, Anushree Agrawal, Andrew Peng, Ian Kivlichan, Mehmet Yatbaz, Madelaine Boyd, Anna-Luisa Brakman, Florencia Leoni Aleman, Henry Head, Molly Lin, Meghan Shah, Chelsea Carlson, Sam Toizer, Ryan Greene, Alison Harmon, Denny Jin, Karolis Kosas, Marie Inuzuka, Peter Bakkum, Barret Zoph, Luke Metz, Jiayi Weng, Randall Lin, Yash Patil, Mianna Chen, Andrew Kondrich, Brydon Eastman, Liam Fedus, John Schulman, Vlad Fomenko, Andrej Karpathy, Aidan Clark, Owen Campbell-Moore