When OpenAI introduced GPT-4o in May 2024, the headline feature wasn't a benchmark score. It was latency. The model could listen to a spoken question and start answering in a few hundred milliseconds, closer to the rhythm of a real conversation than anything a text-based assistant had managed before. The "o" stood for "omni," and the pitch was straightforward: one model, natively handling text, audio, and vision together, instead of stitching together separate speech-to-text and text-to-speech systems around a language model core.

That architectural choice is the part that mattered most, and it's worth explaining plainly for anyone who wasn't following the research at the time.

Why "natively multimodal" was the actual news

Before GPT-4o, a typical voice assistant built on a large language model worked in three hops: a speech-recognition system transcribed your audio to text, the language model reasoned over that text, and a separate text-to-speech system read the answer back. Each hop added delay, and each hop lost information (tone, pauses, emphasis, background sound) that never made it into the transcript the model actually reasoned over.

GPT-4o collapsed that pipeline. A single model was trained end-to-end across text, audio, and image inputs, which is why it could react to *how* something was said, not just what was said, and why it could describe a live video feed in something close to real time. It's a meaningfully different engineering approach from bolting speech modules onto a text model, and it's the reason the demo, a model that could be casually interrupted mid-sentence the way a person can, felt different from earlier voice assistants.

What it actually changed for developers

For people building products on top of the model, the more durable impact was cost and access rather than the voice demo itself. OpenAI priced GPT-4o well below its predecessor while improving benchmark performance across most standard evaluations, and made a lightweight version available for free-tier users. That combination of cheaper, faster, and natively multimodal is what pushed a wave of products to add vision and voice features that would have been impractical to build and serve at the previous generation's pricing.

It also reset expectations across the industry. Within the following year, competing labs each shipped their own natively multimodal or near-real-time voice offerings, a pattern covered in more detail in our explainer on what a large language model actually is, which walks through why "multimodal" became the default expectation rather than a differentiator.

Where it's still catching on

Two years later, the gap between the demo and the deployed reality is narrower than it used to be, but it hasn't closed completely. Real-time voice interaction still degrades in noisy environments, still occasionally talks over a user in ways a human wouldn't, and vision understanding remains stronger for reading and describing static scenes than for tracking fast motion. Enterprise adoption of the voice mode specifically has lagged the more mundane use cases (document analysis, coding assistance, structured data extraction) where GPT-4o's speed and multimodal input mattered without needing the conversational polish.

That's a familiar pattern in this industry: the splashiest capability at launch is rarely the one that ends up doing the most work a year later. For a closer look at how a related architectural bet played out, see our piece on Gemini's long-context strategy, which made a different trade-off: prioritizing how much a model could read over how fast it could talk.

The lasting effect

What GPT-4o actually left behind isn't a single feature. It's a baseline. "Natively multimodal, low-latency, and priced for high-volume use" is now the bar new frontier models are measured against, not an optional add-on. That's a fairly unglamorous legacy for a launch that was, at the time, one of the more visually striking AI demos to date. But it's the kind of shift that's easy to underrate in the moment and obvious in hindsight, which is usually a sign it was the real story.