GPT Models Deprecation

GPT Models Deprecation

From this day forward, all premium API clients will be able to avail the capabilities of GPT-4. OpenAI introduced the ChatGPT API in March 2023, and earlier this month they rolled out their first set of improvements to these chat-centric models. OpenAI envisions a future where chat-oriented models can cater to any requirement. As part of this progression, they are currently announcing GPT Models Deprecation – a strategy to phase out previous versions of the Completions API, urging users to transition to the Chat Completions API.

GPT-4 API general availability

GPT-4 stands as the most proficient model. Since its introduction in March, millions of developers have actively using the GPT-4 API. The portfolio of creative products utilizing GPT-4 continues to expand each day. As of today, all existing API developers with a successful payment history can gain access to the GPT-4 API with an 8K context. OpenAI aims to extend access to new developers by the end of the month. The company also plan to incrementally increase rate-limits thereafter, contingent on the availability of computing resources.

OpenAI is also launching the GPT-3.5 Turbo, DALL·E, and Whisper APIs for general use. The organization is actively working towards the safe facilitation of fine-tuning for GPT-4 and GPT-3.5 Turbo, and anticipates this feature to be available later this year.

Moving from text completions to chat completions

OpenAI unveiled the Chat Completions API in March, and it now constitutes a whopping 97% of their API GPT usage.

The initial Completions API, launched in June 2020, offered a freeform text prompt for engagement with their language models. OpenAI has since discovered that they can often deliver superior results with a more organized prompt interface. The chat-centric approach has demonstrated its robustness, accommodating a vast array of previous applications as well as new conversation-centric requirements. It also offeris greater flexibility and precision.

The structured interface of the Chat Completions API – system messages and function calling. The feature empower developers to create conversational experiences and perform a wide spectrum of completion tasks. This also strengthens security by reducing the risk of prompt injection attacks. It also allows systematic separation of user-provided content from instructions.

Deprecation in the Completions API

In line with their intensified focus on the Chat Completions API and their drive to maximize their computational capacity, OpenAI will be decommissioning some of their older models that use the Completions API in six months. While this API will continue to be available, it will henceforth be tagged as “legacy” in their developer documentation. OpenAI intends to concentrate future model enhancements on the Chat Completions API and does not anticipate releasing new models using the Completions API publicly.

Beginning on January 4, 2024, older completion models will be phased out and replaced with the following models:

Older modelNew model

Applications that use the stable model names for the base GPT-3 models (ada, babbage, curie, davinci) will be automatically updated to the newly listed models on January 4, 2024. These new models will also be ready for preliminary testing in the upcoming weeks by denoting the following model names in API calls: ada-002, babbage-002, curie-002, davinci-002.

Developers utilizing other older completion models (like text-davinci-003) will have to manually update their integrations by January 4, 2024, by specifying gpt-3.5-turbo-instruct in the “model” parameter of their API requests. This new model, gpt-3.5-turbo-instruct, is an InstructGPT-style model, trained similarly to text-davinci-003. It is a direct substitute in the Completions API and will be available for early testing in the forthcoming weeks.

Developers who wish to continue employing their fine-tuned models after January 4, 2024, will need to fine-tune their replacements using the new base GPT-3 models (ada-002, babbage-002, curie-002, davinci-002), or newer models (gpt-3.5-turbo, gpt-4). Once this feature is available later this year, OpenAI will prioritize access to GPT-3.5 Turbo and GPT-4 fine-tuning for users who had previously fine-tuned older models. OpenAI recognizes the difficulty in migrating from models that are fine-tuned on proprietary data. To help with this, they plan to provide assistance to users who have previously fine-tuned models. This will be done to ensure the transition process is as smooth as possible.

In the following weeks, OpenAI will contact developers who have recently used these older models and will share additional information.

Deprecation of embeddings models

Users of older embeddings models, such as text-search-davinci-doc-001, will need to transition to text-embedding-ada-002 by January 4, 2024. OpenAI launched text-embedding-ada-002 in December 2022, and it has proven to be more capable and cost-effective than its predecessors. As of today, text-embedding-ada-002 makes up a staggering 99.9% of all embedding API usage.

OpenAI acknowledges that this represents a significant shift for developers who are utilizing the older models. OpenAI has committed to covering the financial cost associated with users re-embedding their content using these new models. They will be reaching out to affected users in the coming days to provide further information and assistance.

Older modelNew model

Deprecation of the Edits API

Users of the Edits API and its corresponding models, like text-davinci-edit-001 or code-davinci-edit-001, will need to transition to GPT-3.5 Turbo by January 4, 2024. The Edits API beta was an exploratory initiative aimed at enabling developers to receive an edited version of the prompt in line with provided instructions. The feedback received from the Edits API was integral to the development of gpt-3.5-turbo and the Chat Completions API. These can now be utilized for the same purpose.

Read more related articles: