I’m a little late on this one — sorry if you’ve already seen it — but OpenAI just dropped something that I think is going to be a much bigger deal than the headlines suggest.
GPT-4-0613 is now available, and with it comes a new capability: function calling. The short version? You can now describe functions in your prompts, and the model will intelligently choose to output a JSON object containing the arguments needed to call those functions based on user input.
If that sounds dry, stay with me. This is one of those infrastructure moments that changes everything downstream.
What Function Calling Actually Means
Up until now, the way most people interact with GPT-4 is conversational. You ask it something, it gives you text back. Powerful, sure — but it’s fundamentally a text-in, text-out loop. If you wanted GPT-4 to actually DO something — hit an API, query a database, trigger a workflow — you had to build a bunch of fragile middleware to parse its text output and hope it stayed in the format you expected.
Function calling changes that. Now you define a set of functions — with parameters, descriptions, the works — and the model can decide on its own when to invoke one. It returns clean, structured JSON that’s ready to plug directly into your code.
This is the difference between an AI that talks about doing things and an AI that actually does them.
Why This Is the Real AI Integration Layer
I’ve been thinking about knowledge ingestion and the practical, buildable side of AI for a while now. If you’ve been following along, you know my take: the REAL value of these models isn’t in chatting. It’s in connecting them to systems, data, and workflows.
Function calling is the bridge. Think about what becomes possible:
- A GPT-4 powered interface that can query your CRM, pull a customer record, and draft a response — all in one interaction
- An AI assistant that doesn’t just tell you what to do with your calendar, but actually books the meeting
- Natural language interfaces to ANY API, where the model figures out which endpoint to call and what parameters to pass
This is what I was getting at when I was talking about knowledge ingestion being the killer app. The models are smart enough. The bottleneck has been integration — getting them to reliably interact with structured systems. Function calling solves that in a way that’s elegant and, frankly, obvious in hindsight.
A Quick PSA While I’m Here
Heads up. I chat with a LOT of people about GPT, and I keep running into the same thing — they’re using GPT-3.5. The free tier. They’ve never paid the $20/month for GPT-4.
BUY GPT-4.
I’m serious. It is an impossibly valuable investment. It’s 10X better than 3.5 — and I don’t say that casually. The reasoning is sharper, the context handling is leagues ahead, and now with function calling it’s becoming a genuinely different tool. If you disagree after trying it, I’ll reimburse you. That’s how confident I am.
Never before has the cost to intelligence been so minute. Twenty dollars a month for what is effectively an on-demand expert across virtually every domain? We’re living in a strange and wonderful moment and most people are sitting it out over the cost of a couple of lunches.
What Builders Should Be Thinking About Right Now
If you’re building anything that touches AI — and at this point, you probably should be — function calling should immediately change your architecture thinking. A few things I’d be focused on:
1. Define your function library. What are the core actions your application needs to take? Map those out as callable functions with clear parameter schemas. This is your new interface layer.
2. Think in terms of orchestration, not conversation. The model isn’t just generating text anymore. It’s making decisions about which tools to use. Your system design needs to account for that — error handling, authentication, rate limiting on the API calls the model triggers.
3. Watch the cost curve. Right now function calling is available on both GPT-4-0613 and GPT-3.5-turbo-0613. For simpler routing tasks, you might not even need GPT-4. But for complex multi-step reasoning about WHICH function to call with WHAT arguments — that’s where the GPT-4 gap shows up.
The Bigger Picture
I keep coming back to the same theme: the raw intelligence of these models has outpaced our ability to connect them to the world. We’ve had this incredibly powerful brain sitting in a box, and we’ve been passing notes back and forth through a slot in the door.
Function calling kicks the door open. Not all the way — we’re still early. But the trajectory is pretty clear. AI models that can reason about actions, choose the right tool for the job, and execute with structured precision? That’s not a chatbot anymore. That’s an agent.
And if you’ve been watching what AutoGPT and similar projects have been trying to do — build autonomous AI that can take multi-step actions — you’ll recognize that function calling gives those efforts a MUCH more reliable foundation to build on.
The gap between “AI that talks” and “AI that does” just got a lot smaller. If you’re not building with this yet, now’s the time to start.
