OpenAI opened its ChatGPT and Whisper APIs on 1 March, letting developers add text generation and speech recognition to their own applications.
Sulat illustration. Logo: OpenAI.
The ChatGPT model, gpt-3.5-turbo, costs US$0.002 per 1,000 tokens. Hosted Whisper large-v2 costs US$0.006 per minute, with endpoints for transcription in the original language and translation into English.
Developers can select the moving gpt-3.5-turbo alias or pin the gpt-3.5-turbo-0301 snapshot, supported through at least 1 June.
API data is excluded from model training unless organisations opt in. The default retention period is 30 days, with stricter options available.
A conversation in messages
The launch-day developer guide describes two required request fields: a model name and a list of messages. Each message carries a role and content. System messages establish instructions, user messages supply requests and assistant messages represent replies.
Applications can include earlier exchanges in that list. The response contains a generated assistant message and usage counts for the prompt, completion and total tokens. A single instruction can also use this format without a running conversation.
Speech recognition without local hosting
OpenAI released Whisper’s models and inference code in September 2022. The system was trained on 680,000 hours of multilingual audio. Its encoder-decoder architecture processes audio chunks and predicts text, with task markers distinguishing transcription from translation. The hosted service gives application developers another route to that speech-processing technology.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
