OpenAI confirmed on Friday that it will hold off on a broad public launch of its GPT‑5.6 AI models after the White House asked the company to pause the release. The request comes just two weeks after the administration forced Anthropic to take its most advanced models offline, creating a tense backdrop for frontier AI labs.

Instead of a full rollout, OpenAI will initially provide GPT‑5.6 to a limited set of customers that have been preapproved by the U.S. government. The company said it will submit a list of prospective users to the White House, receive feedback, and then grant access only to those cleared by officials. OpenAI plans to broaden the pool of approved customers next week, including some international partners, though details of the approval process remain confidential.

In a blog post, OpenAI called the delay a "short‑term step" taken to protect broader availability in the weeks ahead. The firm emphasized that it does not view government vetting as a permanent model for future releases, arguing that prolonged oversight would keep powerful tools from developers, enterprises and cybersecurity defenders who need them.

The request aligns with a recent executive order signed by President Trump aimed at addressing cybersecurity risks posed by advanced AI systems. The order calls for a "voluntary process" that would let AI labs share models with the government thirty days before a wider launch, while explicitly stating the process should not become a de facto licensing regime. OpenAI said no such voluntary framework exists yet, leaving the industry in a gray area where cooperation with the government is encouraged but not yet formalized.

Anthropic’s earlier compliance with a White House export‑control directive, which led the company to shut down its most advanced models for all customers, remains unresolved. The overlap of these two high‑profile cases underscores growing uncertainty for U.S. AI developers as regulators balance innovation against national security concerns.

OpenAI described GPT‑5.6 as arriving in three flavors. "Sol" represents the most capable version, topping benchmarks in cybersecurity, biology and agentic tasks. "Terra" offers a middle‑tier balance of performance and cost, while "Luna" is positioned as a fast, affordable option for broader use cases. All three variants incorporate a "layered safeguard stack" designed to block malicious actors from exploiting the models for cyberattacks or other harmful activities.

Company executives declined to elaborate on how the safeguard stack works, noting that specific technical details are proprietary. Nevertheless, they stressed that the safeguards are a core part of the model’s design and will be continuously updated as new threats emerge.

OpenAI hopes to expand access to GPT‑5.6 within the coming weeks, pending further collaboration with the administration to shape a repeatable process for future model releases. The firm’s leadership reiterated its commitment to making the technology widely available while navigating the evolving regulatory landscape.

Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.