One Year Into the GPT-3 API: What Are Developers Actually Building?

It’s been roughly a year since OpenAI opened up commercial API access to GPT-3, and I’ve been trying to keep a running list of what people are actually shipping with it, as opposed to what people are merely tweeting demos of. The two categories overlap less than you’d think.

The biggest, most obvious category is copywriting assistants. Feed the model a product description and a tone, and it drafts ad copy, product blurbs, email subject lines, that sort of thing. This makes sense as a first wave — marketing copy is short, low-stakes if it’s slightly off, and easy for a human to skim and approve or discard. Several small startups have built entire products around this single use case: you type a few words about what you’re selling, and you get back five or six variations to choose from. It’s not glamorous, but it’s the kind of task where “good enough, fast” beats “perfect, slow,” and that’s exactly GPT-3’s sweet spot right now.

Chatbots are the second big cluster, though “chatbot” undersells what’s happening. These aren’t the brittle decision-tree bots of a few years ago. Because GPT-3 can hold a conversational thread and respond in whatever voice you prompt it with, people are building everything from customer-support front-ends to characters you can just talk to for fun. The quality is wildly inconsistent — sometimes it’s uncannily coherent, sometimes it confidently says something false or nonsensical — but the fact that a small team can stand up something conversational in a weekend, without touching a training pipeline, is new.

Code helpers are the interesting one

The third cluster is smaller but, to me, the most interesting: code-helper prototypes. People are experimenting with prompting GPT-3 to translate plain-English descriptions into snippets of code, or to explain what a block of code does, or to autocomplete boilerplate. None of what I’ve seen is production-ready — think toy demos and weekend hacks rather than anything you’d trust in a real codebase — but the direction is obvious enough that it’s hard to ignore. If a general-purpose language model can already stumble its way through generating a SQL query or a regex from a sentence, it’s not a huge leap to imagine tools built specifically for code getting quite good at this within the next couple of years.

What ties all of this together is that almost none of these projects required their builders to know anything about machine learning. You get an API key, you write a prompt, you parse the text that comes back. The barrier to entry has dropped from “hire a research team” to “learn to write a good prompt,” and that’s the real story here — not any single flashy demo, but the sheer number of small teams and solo developers now able to build language-based products at all.

The obvious caveats still apply: costs add up at scale, the API has usage limits and a waitlist for broader access, and the model’s tendency to occasionally output plausible-sounding nonsense makes it a poor fit for anything where accuracy really matters. But a year in, this already looks less like a one-off research curiosity and more like the beginning of a genuine new tool category. I’ll be curious to see, a year from now, whether these prototypes have turned into real products or quietly faded out.