OpenAI has lifted the curtain on GPT-5.6, a new family of models promising advancements across several areas, from language understanding to cybersecurity. The news, in its brevity, signals something deeper: the company is pushing hard on the performance-security combination, but always within a cloud paradigm that leaves a whole segment of the enterprise market uncovered.

The emphasis on cybersecurity is anything but accidental. Enterprises, especially those in regulated industries, are increasingly reluctant to send sensitive data to an external API, no matter how robust. An LLM like GPT-5.6 may excel at detecting phishing attempts or fixing code vulnerabilities, but if using it requires shipping my source code outside the corporate perimeter, the residual risk isn't technical — it's governance. OpenAI knows this, and the announced improvements seem designed to reassure CISOs, not to change the deployment architecture.

Who gains in the short term are companies with mature security processes and low-criticality data, which can integrate the model via API and get an almost immediate boost. Who loses is the galaxy of organizations — banks, defense, healthcare, public administration — for which data residency is non-negotiable. For them, a more capable model inaccessible on-premise is like a supercomputer in a locked room: technically interesting, operationally irrelevant.

OpenAI's choice signals a structural trajectory. The major providers of foundational models are competing on brute power and perceived trust, not on deployment flexibility. This consolidates a two-speed ecosystem: on one side, cloud models, increasingly large and vertically integrated; on the other, a growing demand for self-hosted LLMs, driven by TCO, latency, and sovereignty needs. GPT-5.6, with its cybersecurity improvements, makes the gap even more apparent: those who can use it in the cloud will gain a competitive edge, those who cannot will have to look elsewhere, accelerating the maturation of open-source alternatives and frameworks like vLLM or Ollama.

For technical decision-makers, the question is not whether GPT-5.6 is more secure than its predecessor, but whether the promised security compensates for ceding control over data. In this sense, the new model family doesn't rewrite the rules of enterprise adoption, but sharpens them: the AI game in the enterprise is increasingly played on the physical boundary of the infrastructure. For those evaluating on-premise deployment, complex trade-offs between performance, cost, and innovation pace exist, and AI-RADAR offers analytical frameworks to map them without shortcuts.