OpenAI Slashes GPT-3.5 API Costs by 90%, Whisper Goes Public

OpenAI just flipped the switch on a major API expansion, making the GPT-3.5 Turbo and Whisper models broadly available to developers. If you’ve been waiting for the pricing to make sense, your moment has arrived. Through what the company calls ‘system-wide optimizations,’ they’ve driven costs down by 90% since December and are passing the savings on. The new gpt-3.5-turbo model is priced at $0.002 per 1,000 tokens—a full order of magnitude cheaper than the existing GPT-3.5 models. That’s not a typo.

But the real story here isn’t just the price cut. It’s the format shift. Unlike traditional GPT models that chew on unstructured text, this new API endpoint is built for a sequence of messages and metadata, using a new format called Chat Markup Language (ChatML). That means early testers are already migrating from text-davinci-003 to gpt-3.5-turbo with minimal prompt adjustments. The company is also giving developers more control over model versions. The gpt-3.5-turbo alias will always point to OpenAI’s recommended stable build, but if you need to lock things down, you can specifically call gpt-3.5-turbo-0301, which is guaranteed support through at least June 1st.

Several big-name partners are already all-in, and their use cases show the range here. Snap introduced ‘My AI’ for Snapchat+ subscribers this week, putting a customizable chatbot in front of 750 million monthly users. Quizlet, which has been working with OpenAI for three years, is launching Q-Chat—an adaptive AI tutor that turns study materials into a conversational experience for its 60 million students. Instacart plans to launch ‘Ask Instacart’ later this year, combining GPT-3.5 with product data from over 75,000 retail locations to answer open-ended food questions with shoppable results. And Shop, Shopify’s consumer app with 100 million shoppers, is using the API to power a new shopping assistant that scans millions of products for personalized recommendations.

Then there’s Whisper. The open-source large-v2 speech-to-text model is now available through the API with a speed and cost-efficiency upgrade. Speak, a language learning app that’s the fastest-growing English platform in South Korea, is already using it to power an AI speaking companion. They’re betting on Whisper’s ability to handle accented, imperfect speech from learners—something that would make traditional speech-to-text systems choke. The company is expanding that product globally on the back of this API. If you’re building anything that involves voice, the barrier to entry just collapsed.