GPT-4 and What It Means for Software Builders

GPT-4 arrived on March 14th, 2023. I got access to the API a week early through OpenAI's developer program. After two months of ChatGPT on GPT-3.5, the jump to GPT-4 isn't incremental. It's noticeable enough that it changes what you can actually build.

I need to be careful not to oversell this. The headlines claiming GPT-4 is AGI or approaching human-level intelligence are nonsense. But if you're actually building software, the practical differences matter. The model is better at reasoning, better at handling edge cases, better at understanding context. It hallucinates less. It's more reliable in production.

The question for builders like me is: what does this actually unlock?

Where GPT-4 Actually Improves Over 3.5

Let me be specific. GPT-3.5 is phenomenal at generating text that sounds good. It can write an email, a blog post, code that compiles. But ask it to reason through a problem step-by-step, and it often takes shortcuts. It'll miss edge cases. It'll make leaps in logic that don't hold up.

GPT-4 is noticeably better at this. I tested it on the same set of tasks I'd been experimenting with using 3.5. Parsing complex instructions? Better. Understanding multi-step workflows? Better. Catching logical inconsistencies in code? Better. Not perfect—it still makes mistakes—but the error rate is lower.

The consistency is what matters to me as someone building products. With 3.5, you'd write a prompt, test it 10 times, and get 8 good outputs and 2 bad ones. With 4, the ratio improves to 9 good, 1 bad. In production, that matters. That's the difference between needing heavy human review and needing light review.

For coding tasks, the improvement is tangible. GPT-4 can write longer code blocks without losing context. It understands architectural patterns better. I had it write a middleware authentication service that actually worked on the first attempt. With 3.5, I would've had to debug and iterate.

But There Are Real Trade-offs

Here's what people don't talk about: GPT-4 is slower. API calls take 3-5x longer than 3.5. In February when I was prototyping real-time features, that latency was acceptable. In March with GPT-4, it's noticeable. Users will wait 2-3 seconds for a response. Some features that felt snappy with 3.5 feel sluggish with 4.

It's also more expensive. GPT-4 costs 15x more for input tokens and 30x more for output tokens compared to 3.5. For an application processing thousands of requests daily, that difference compounds quickly.

What this means: GPT-4 isn't a drop-in replacement for 3.5 in your existing product. You can't just swap the model and call it a day. You need to redesign around the latency. You need to calculate whether the accuracy improvement justifies the cost increase. For some use cases, it's a no-brainer. For others, 3.5 is still the right call.

What This Opens Up

The reliability improvement does unlock new applications. The most interesting one to me: ERP workflow automation. I'm imagining an AI assistant that handles data entry in your system. With GPT-3.5, you'd need heavy guardrails and review. The model would make too many mistakes. With GPT-4, the error rate is low enough that a human spot-check once a week might be sufficient.

Another possibility: document understanding. Contracts, invoices, complex forms. GPT-3.5 could extract basic fields. GPT-4 can understand nuance—flagging unusual clauses, summarizing terms, highlighting risks. That's useful in ERP systems where a lot of work is document processing.

Natural language interfaces are more viable now. GPT-3.5 would misinterpret instructions too often. GPT-4 is reliable enough that you could let users query their ERP data using plain English and trust the outputs most of the time.

The Implications for Product Strategy

Here's what I'm thinking about as a builder: the moat around AI products just got thinner. If your entire competitive advantage is "we use AI," and you can swap GPT-3.5 for GPT-4, then your advantage is temporary. What matters now is how you integrate the model into your specific product. The workflow design. The data preparation. The human review layers you build around it.

For Arvension, this means focusing on the business logic layer, not the AI layer. We can build a better ERP system that uses GPT-4 for specific workflows—but the value isn't the AI, it's the system that wraps around the AI and makes it useful for operations teams.

The other implication: costs matter more now. With 3.5, the computational cost of AI features was almost negligible. With 4, you need to think about unit economics. Is your user willing to pay for the extra processing cost of GPT-4, or do you need to absorb it? That's a business decision, not a technical one.

What I'm Building Next

For my current projects, I'm running a parallel experiment: some workflows on GPT-3.5, some on GPT-4. Measuring accuracy, latency, and cost for each. My guess is that the sweet spot is 80% of workflows run on 3.5 for speed and cost, and 20% run on 4 for the high-judgment tasks where accuracy matters more than speed.

GPT-4 is a meaningful improvement. It's not revolutionary. But for someone building actual products, it opens doors that weren't viable three months ago. That's worth paying attention to.

The AI landscape is moving fast. By the end of 2023, there'll be more models, more capable, more specialized. But right now, in March, GPT-4 is the model to optimize for if you're serious about integrating AI into production systems.