OpenAI DevDay London: o1 Coding and Realtime Voice Demos

OpenAI DevDay London has just taken place, and I want to show you several demonstrations. First, the team presents a full-stack drone-control application whose code is attributed to o1. In the clip, a person starts the server, checks the battery and video stream, then tests takeoff and the interface buttons. The video shows a human operating the interface, not o1 independently flying the drone.

The video includes sponsorship from the AI Port and Turing Post newsletters. I mention their project comparing how image models depict generations and discuss how model outputs can reflect stereotypes; those examples illustrate the project rather than establish a universal finding about every model or generation.

Next is Wanderlust, a voice-enabled travel application. I describe this as o1 using voice, but OpenAI’s contemporary documentation identifies the Realtime API audio model as GPT-4o realtime preview. The demo shows an application integrating that API with its own tools; it does not establish that o1 powered the voice interaction.

The presenter asks about a trip to Singapore, places to visit and hotels. As the conversation proceeds, the app displays a transcript, updates a map and brings up hotel details. The presenter explains that the application can connect the Realtime API to developer-defined functions, APIs and tools.

The tool integration lets the app place map pins, change the view and display information. This is a demonstration of an application built around voice and tool calls; its success in this staged setting is not a guarantee that real travel information or every action will be correct.

The presenter changes the setting back to London near The Brewery on Chiswell Street and asks for nearby pie shops for the audience. The app displays three options. The presenter explicitly notes that the second, Ilan’s Posh Pies, is a fictional demo location, even though the other two are real.

To extend the demonstration beyond the interface, the team connects the Realtime API to Twilio phone calling. A colleague plays the shop representative. The presenter asks the assistant to arrange about 200 pies, split between meat and vegetarian options, and keep the budget under £1,000.

During the staged call, the assistant asks about available pies and the price of 100 meat and 100 vegetarian pies. The shop initially quotes about £1,200, then agrees to £1,000 in exchange for being featured at the event. That reaches the stated £1,000 limit rather than falling strictly below it.

The call continues to a delivery arrangement near the venue. The exchange illustrates how an application can connect a spoken assistant to a telephone service and a human counterpart; it is a scripted demo setting with a fictional business, not evidence of a real commercial order.

A second demonstration uses a voice interface to explore the solar system. The presenter asks for the largest planet, and the assistant identifies Jupiter. The conversation then moves to Earth and its characteristics, while the interface changes what is displayed.

Asked about the balance of land and water, the assistant gives approximately 29% land and 71% water and displays a pie chart. The presenter then asks about Mars and its volcanoes, prompting a response about Olympus Mons and a comparison with other large volcanoes.

I see this kind of responsive dialogue and visualization as a promising direction for education. I describe the content as dynamically generated rather than predefined, although the clip alone does not establish which parts were generated or prebuilt. Educational answers and graphics should still be checked for accuracy.

OpenAI also announced five new voices and prompt-caching discounts: at that time, cached text inputs were discounted 50% and cached audio inputs 80%. These figures describe the October 2024 update, not current rates; OpenAI’s later documentation should be consulted for current pricing.

Finally, I report that DevDay audience members were offered o1-preview, while OpenAI product leader Olivier Godement said the full o1 was still being developed. My suggestion that the audience access might reflect European availability was speculation. OpenAI later released an improved o1, and its current o1 page says o1-preview and o1-mini have been retired from the API and o1 is deprecated.