OpenAI drops o1 API, slashes realtime audio prices by 60%

OpenAI just made its most capable reasoning model, o1, available to developers on usage tier 5, packing in production-ready features that were conspicuously absent from the o1-preview. Function calling, Structured Outputs, developer messages, and vision capabilities are all live now. This isn’t just a chatbot upgrade — it’s OpenAI signaling that o1 is ready to power agentic workflows in customer support, supply chain optimization, and financial forecasting. The new snapshot, o1-2024-12-17, uses 60% fewer reasoning tokens on average than its preview predecessor, which directly translates to lower latency and cheaper API calls. A new reasoning_effort parameter also lets developers dial in how much compute the model burns before answering.

On the realtime front, the changes are just as aggressive. OpenAI integrated WebRTC support directly into the Realtime API, abstracting away the nightmare of audio encoding, noise suppression, and congestion control. You can now spin up a low-latency voice experience with a few lines of JavaScript. That technical simplification comes alongside a brutal price cut: GPT-4o audio tokens are now 60% cheaper, with cached audio input dropping by 87.5% to just $2.50 per million tokens. GPT-4o mini is also joining the Realtime API beta at one-tenth the previous audio rates, making real-time voice economically viable for apps that couldn’t justify the cost before.

OpenAI also rolled out Preference Fine-Tuning, a new customization method that steers models based on user and developer preferences rather than just raw examples. It’s a more nuanced approach than standard supervised fine-tuning and suggests OpenAI is serious about letting enterprises bake their own policies and tone into models. New Go and Java SDKs, both in beta, round out the developer push. The company is clearly trying to make its stack stickier for engineering teams that aren’t all-in on Python.

The o1 model itself sets new state-of-the-art results on several benchmarks, and OpenAI claims it significantly outperforms GPT-4o on function calling and structured output tests. But access is still gated behind usage tier 5 and rolling out incrementally. The version shipping today is the same post-trained model that hit ChatGPT two weeks ago, tuned further based on behavioral feedback. If you’re not on tier 5 yet, you’re still watching from the sidelines — though the pace of these releases suggests the gates won’t stay up for long.