OpenAI launches the ChatGPT and Whisper APIs

Background

ChatGPT had reached about 100 million monthly users two months after its launch, and all of that traffic ran through OpenAI's own chat product^[2]. OpenAI already sold access to GPT-3.5 models through its API, and the new model was priced against exactly those existing models^[1].

What was announced

On 1 March 2023 OpenAI made gpt-3.5-turbo, the model behind ChatGPT, available through its API^[1][2]. On the same day it opened both the ChatGPT and Whisper APIs to developers^[4]. The Whisper API handled a different job from the chat model: speech-to-text transcription and translation^[3]. One API took text in and gave text back; the other took audio in.

The price

gpt-3.5-turbo cost $0.002 per 1,000 tokens, a price OpenAI said was 10 times cheaper than its existing GPT-3.5 models^[1]. A token is the unit these APIs bill in: text is broken into pieces — a short word, part of a long word, a punctuation mark — and the charge is counted per piece, so the bill grows with the amount of text sent in and returned. Quoting a price per thousand tokens lets a developer estimate what a feature will cost before building it, which matters more for a product used by many people than for a demo.

The Decoder reported that a series of optimizations had reduced ChatGPT's cost by 90 percent since December 2022^[4]. Serving a model is the expensive part of running it as a product, so a fall in that cost and a tenfold cut in the developer price point in the same direction: the same model became cheaper to run, and the price list was rewritten around that.

The first customers

Instacart, Snap and Quizlet were among the first customers^[2]. Companies that already have an app and an audience are the natural early buyers of a per-token API: they can add a language-model feature to something people already use without training or hosting a model of their own. The 100 million monthly users ChatGPT had gathered in its first two months were evidence that the demand existed^[2]; the API was a way to sell the same capability to businesses that wanted it inside their own products rather than inside OpenAI's chat window.

Why it mattered

Three things changed for developers that day. The model that had made ChatGPT notable could be called from their own code, so a chat-like feature no longer required building the model itself^[1][2]. The price was published per thousand tokens and set ten times below the existing GPT-3.5 models, which made the running cost of a feature something a team could calculate in advance^[1]. And speech could be handled by the same provider, so a product that needed to turn audio into text did not have to bring in a second vendor^[3].

Publishing a model through an API also separates the model from the interface. The same gpt-3.5-turbo could appear inside OpenAI's own chat product and inside someone else's application at the same time, with usage billed by the token rather than by seat or by month. That is a different way of distributing a model from shipping a single website, and it is the form in which most software builders first got their hands on a ChatGPT-class model^[1][2].