GPT-4 API General Availability and Completions API Deprecation
OpenAI has announced the general availability of the GPT-4 API for all paying customers and a comprehensive deprecation schedule for legacy models within the Completions, Embeddings, and Edits APIs. This shift marks a strategic move toward the Chat Completions API as the primary interface for developer interactions with OpenAI's most capable models.
GPT-4 API General Availability
All existing API developers with a history of successful payments now have access to the GPT-4 API with an 8K context window. OpenAI plans to extend access to new developers by the end of April 2024, with rate limits to be increased based on compute availability.
In addition to GPT-4, the GPT-3.5 Turbo, DALL·E, and Whisper APIs are now generally available. OpenAI also stated that fine-tuning for GPT-4 and GPT-3.5 Turbo is expected to be released later in 2024.
Transition to Chat Completions API
OpenAI is prioritizing the Chat Completions API, which currently accounts for 97% of its API GPT usage. The company is moving away from the original freeform text prompt interface of the Completions API in favor of a structured prompt interface.
Advantages of the Chat-Based Paradigm
- Enhanced Flexibility: The structured interface supports system messages and function calling, enabling more complex conversational experiences and completion tasks.
- Improved Security: The Chat Completions API reduces the risk of prompt injection attacks by structurally separating user-provided content from system instructions.
- Multi-turn Capabilities: The API is designed specifically for multi-turn conversations, providing higher specificity in responses.
OpenAI is working to address remaining gaps in the Chat Completions API, including increasing steerability to reduce "chattiness" and providing log probabilities for completion tokens.
Deprecation of Legacy Models
OpenAI will retire several older models across three different API categories. The Completions API will be labeled as "legacy" in developer documentation, and no new models will be publicly released for it.
Completions API Deprecation
Starting January 4, 2024, older completion models will be replaced by the following:
| Older Model | New Model |
|---|---|
| ada, babbage | babbage-002 |
| curie, davinci | davinci-002 |
| davinci-instruct-beta, curie-instruct-beta, text-ada-001, text-babbage-001, text-curie-001, text-davinci-001, text-davinci-002, text-davinci-003 | gpt-3.5-turbo-instruct |
Applications using stable model names for base GPT-3 models (ada, babbage, curie, davinci) will be automatically upgraded. Users of other models, such as text-davinci-003, must manually upgrade to gpt-3.5-turbo-instruct.
Embeddings API Deprecation
All users of older embeddings models (e.g., text-search-davinci-doc-001) must migrate to text-embedding-ada-002 by January 4, 2024. OpenAI noted that text-embedding-ada-002 currently accounts for 99.9% of all embedding API usage.
Edits API Deprecation
Users of the Edits API and its associated models (e.g., text-davinci-edit-001 or code-davinci-edit-001) must migrate to GPT-3.5 Turbo by January 4, 2024.
Fine-Tuning Migration Support
Developers who have fine-tuned models on older base GPT-3 models must fine-tune replacements on babbage-002, davinci-002, GPT-3.5 Turbo, or GPT-4 to continue using them beyond January 4, 2024. OpenAI has committed to providing priority access to GPT-3.5 Turbo and GPT-4 fine-tuning for users who previously fine-tuned older models and will cover the financial cost of users re-embedding content for the Embeddings API migration.