Separating capability from hype in OpenAI's latest releases
Every few months, OpenAI releases a new model that the tech press calls "revolutionary." For enterprise development teams, the question is simpler: does this change what we can build for our clients?
We've been integrating large language models into production systems since GPT-3.5. Here's our honest assessment of what's genuinely useful and what's still marketing.
What actually matters for enterprise applications
Function calling and structured output
This is the most underappreciated feature for enterprise developers. Models can now return structured JSON that matches a predefined schema, and they can "call" functions in your application code. This transforms LLMs from text generators into components of real software systems.
We use this extensively in our agentic AI projects. Instead of parsing natural language output and hoping it's formatted correctly, we define schemas for agent actions and get reliable, structured responses. Error rates dropped from roughly 15% to under 2% when we switched from text parsing to function calling.
Multimodal input
GPT-4o processes images, audio, and text in a single model. For enterprise applications, this means document processing without OCR preprocessing, visual inspection for quality control, and audio transcription integrated with analysis — all in one pass.
Reasoning models
These models think through problems step-by-step before answering. For complex analytical tasks — legal research, financial modeling, code review — they produce meaningfully better results. They're also slower and more expensive, so reserve them for high-value tasks where accuracy matters more than speed.
What's still not enterprise-ready
Autonomous code generation. AI can write code snippets, but generating production-quality software remains unreliable. We use AI coding assistants to accelerate our developers, not replace them.
Unsupervised decision-making. For any decision with regulatory, financial, or safety implications, AI should prepare recommendations for human review, not act autonomously.
How we use these capabilities at Globalbit
Our legal AI system (Psika.ai) uses function calling to structure precedent searches. Our document processing tools use multimodal input to handle mixed Hebrew-English documents. Our code review tools use reasoning models to catch logical errors that simpler models miss.
The common pattern: we use the latest model capabilities where they provide clear value and fall back to deterministic code for everything else.
Frequently asked questions
Should we wait for better models before starting an AI project? No. The architectural patterns and integration work you do today transfer to future models. When a better model arrives, you swap the model layer — the rest of your system stays the same.
Which model should we use? It depends on the task. GPT-4o for general-purpose work, reasoning models for complex analysis, and smaller models for high-volume, lower-complexity tasks. Most enterprise systems use multiple models.
If you're evaluating how AI fits into your enterprise systems, we can walk you through the practical considerations.