In the talk “The New Code – Specs: Write once, run everywhere” (July 2025), Sean Grove from the OpenAI technical team shared the vision that “executable specifications” will become the fundamental unit of software development, shifting the focus from writing code to defining rigorous specifications that serve as the source of truth.

In this approach, a versioned and precise specification can be “compiled” into documentation, test cases, AI model behavior, and even executable code. The idea is to align humans and machines through clear specifications that express intent, rather than focusing everything on source code.

And although Grove presents this vision as a new truth, this is something GeneXus anticipated, applied, and surpassed more than 35 years ago.

GeneXus, created in the late 1980s, has been a pioneer in software generation based on models or Knowledge Bases.

In GeneXus, code is a byproduct, not the end goal. This means that the developer defines what is needed (the problem specification) in a Knowledge Base, and the platform handles how to solve it by generating the code automatically.

For decades, GeneXus has enabled the capture of business knowledge in a structured model and the automatic generation of systems from that model, using deterministic code generators.

Essentially, the concept of “executable specification” that Sean Grove refers to has been implemented by GeneXus from the beginning, using a high-level model as the source of truth, from which complete programs for multiple languages and platforms are automatically derived.

From “Separating the What from the How” to “How We Generate the What”

To generate code from the source of truth, OpenAI uses Large Language Models. GeneXus, on the other hand, uses deterministic generators to create software from the model.

While the hype is on the side of generation with Large Language Models, doing so with deterministic generation has two advantages that aren’t widely discussed and that we want to share, because they directly impact team work and project costs:

• Reliable code with no need for human review.
• Cost efficiency in generation.

1. Reliable Code Without Human Review for Mission-Critical Systems

Unlike code generation with LLMs, GeneXus ensures a very high level of reliability for mission-critical systems without the need for manual, line-by-line review.

This is possible thanks to deterministic generation. GeneXus always produces the same, consistent, and correct code from the specifications defined in the Knowledge Base, guaranteeing quality, traceability, and ease of maintenance.

With GeneXus technology, there is no room for random variability or “hallucinations”—something that can happen with generative AI outputs, potentially leading to serious challenges for critical systems. In fact, solutions relying solely on generative AI may be effective in the short term, but present limitations when applied to large-scale critical systems:

• The code produced by AI can vary even for similar requests, affecting consistency.
• There’s a risk that the generated code contains errors or bugs and must be reviewed before use.
• When building complex systems, generative AI tends to leave out important components or necessary details.
• Moreover, this type of AI still faces challenges in optimally structuring and maintaining each company’s proprietary knowledge.
  
  

In mission-critical contexts where error is not an option, GeneXus allows you to generate 100% consistent and correct code. There’s no need for a human to review or correct the generated code, because the formalized knowledge and the type of generator guarantee model fidelity and result quality.

2. Efficiency and Cost: GeneXus vs. Generative AI Models

Another crucial aspect where GeneXus outperforms the generative AI approach is in computational efficiency and operational cost. AI models like GPT-4, among others, require significant resources for training (massive datasets, cutting-edge GPU clusters). They also consume costly infrastructure every time a result is generated. Every line of code produced by a generative model involves a new inference with significant computational cost. If you need to generate code repeatedly or for multiple platforms, those costs quickly add up.

GeneXus, by contrast, uses deterministic software generation, with algorithms that run on a regular PC, producing optimized code without the need for GPUs or expensive servers.

The cost per generation is basically that of running a desktop program, which trends toward zero at scale. This enables a much more efficient approach.

Compared to generative AI models, GeneXus offers a much more economical and efficient approach, with no hidden execution costs or need for AI infrastructure. The investment is focused on properly modeling knowledge; after that, the generation is handled by the machine with a negligible marginal cost.

A Pioneering, Industrial Model Ready to Scale Today

The vision of “executable specifications” that is now gaining traction validates the path GeneXus charted more than three decades ago, already offering a solid, mature, and industrialized solution for generating software from models, with proven efficiency and reliability in thousands of productive systems.

If you find this approach interesting, I invite you to explore more about this convergence of AI and Low-Code at GeneXus Live:
genexus.com/live
.

I also encourage you to dive deeper into these topics through
genexus.com/webinars
, where you’ll find resources designed to help teams and companies make the leap to this new development paradigm.

We look forward to discovering together how GeneXus can revolutionize the way you create software!

You may also be interested in reading:

GeneXus Next: Native Agentic Low-Code Development for Mission-Critical Systems

Creation and Innovation in the Age of AI

Low-Code + Generative AI: Challenges and Opportunities for CIOs