OpenAI just pulled the trigger on two moves that will reshape how developers build with its tech: GPT-4 is now generally available to all paying API customers, and a whole generation of older models is heading for the scrap heap.
The company said Thursday that any developer with a history of successful payments can now access the GPT-4 API with an 8K context window. That ends a months-long waiting game that began in March when millions of developers first requested access. New developers should get the green light by the end of July, with rate limits rising after that as compute capacity allows. Alongside GPT-4, the GPT-3.5 Turbo, DALL·E, and Whisper APIs are also now generally available.
But the bigger news might be what’s getting retired. OpenAI is officially labeling the original Completions API as “legacy” and will shut down several older models on January 4, 2024. The move isn’t surprising — the company revealed that the Chat Completions API, introduced just four months ago in March, already accounts for 97% of its API GPT usage. The structured interface, with its system messages and function calling capabilities, has clearly won. OpenAI argues the chat-based approach also reduces prompt injection risks by structurally separating user input from instructions.
Developers still relying on models like text-davinci-003 will need to manually switch to gpt-3.5-turbo-instruct, a drop-in replacement trained in a similar InstructGPT style. Those using the base GPT-3 models — ada, babbage, curie, and davinci — will be auto-upgraded to new versions. Fine-tuned model users face the biggest lift: they’ll need to rebuild their custom models on top of the new base models or wait for GPT-4 and GPT-3.5 Turbo fine-tuning, which OpenAI says is coming later this year with priority access for those affected. The company is promising direct support to ease the transition, but a six-month runway for a full fine-tuning migration is tight. You can hear the grumbling from machine learning engineers already.
