OpenAI drops ChatGPT and Whisper APIs at 90% discount

For months, developers have been jumping through hoops—scraping web interfaces or burning through budgets with text-davinci-003—just to build ChatGPT into their own tools.

That stopgap era is over. On March 1, OpenAI officially launched APIs for both GPT-3.5 Turbo and its Whisper speech-to-text model.

The headline here is the sticker price. OpenAI says system-wide optimizations cut ChatGPT running costs by 90% since December, and it is passing those savings directly to developers. The new gpt-3.5-turbo endpoint costs $0.002 per 1,000 tokens—making it ten times cheaper than previous GPT-3.5 models.

Slashing prices and solving privacy headaches

Cost was only half the battle. OpenAI also made major concessions to win over skittish enterprise developers.

Chief among them: data submitted through the API will no longer be used to train models by default, unless organizations explicitly opt in. OpenAI also introduced a 30-day default retention policy and clarified that users own both the prompts they submit and the outputs generated.

To grease the wheels, OpenAI eliminated its mandatory pre-launch app review in favor of automated monitoring. For heavy hitters pushing more than 450 million tokens a day, the company rolled out dedicated Azure-hosted instances that give teams pinned model snapshots and the option for longer context windows.

Whisper is also getting its moment. While the open-source speech model won rave reviews last September, running large-v2 locally was notoriously heavy on hardware. The new Whisper API charges a flat $0.006 per minute for transcription and English translation, handling messy audio without needing a dedicated GPU farm.

Under the hood, developers will need to tweak how they format prompts. Instead of feeding in raw unstructured text, gpt-3.5-turbo takes a structured sequence of messages via ChatML. The initial snapshot, gpt-3.5-turbo-0301, will be supported through at least June 1.

Big names are already moving fast. Snap tapped the API for its new My AI chatbot for Snapchat+ subscribers, Quizlet built an adaptive tutor called Q-Chat, Instacart plans to launch an in-app shopping assistant, and language app Speak plugged Whisper into its oral coaching tool.

OpenAI did add a rare dose of humble pie, admitting its API uptime over the past two months fell short of expectations and promising engineering focus on rock-solid stability.

Why it matters

At $0.002 per thousand tokens, the economics of generative AI completely flipped overnight. Applications that felt like expensive toys—automated customer support, interactive study aids, smart assistants—suddenly have realistic unit economics. You can now process roughly 750,000 words for two bucks.

For the rest of the AI industry, OpenAI just set a brutal baseline. Rival model providers cannot simply match ChatGPT's output quality; they now have to match a cost structure subsidized by massive engineering efficiencies and Microsoft's Azure infrastructure.

Surrendering training rights on API data looks like a tactical retreat, but it is actually OpenAI's smartest play. By removing the fear of corporate data leaking into future model weights, OpenAI cleared the single biggest obstacle keeping conservative enterprise budgets on the sidelines.