OpenAI said Friday it is limiting the debut of its GPT‑5.6 suite to a "small group of trusted partners" after a direct request from the U.S. government. The decision comes under the Trump administration’s recent executive order that obliges leading AI developers to submit their most advanced models for a government review up to 30 days before public launch.
The GPT‑5.6 family comprises three models: Sol, the most powerful offering; Terra, a balanced version intended for everyday tasks; and Luna, a faster, lower‑cost alternative. While Sol boasts new agentic capabilities in coding, biology and cybersecurity, all three models are currently barred from broader release. OpenAI clarified that the limited preview will be shared only with partners whose participation has been disclosed to the administration.
OpenAI’s move mirrors a recent incident involving Anthropic, which was forced to withdraw its flagship Fable 5 model after the government ordered the company to block access for foreign nationals. The Anthropic case sparked a debate over how much authority the government should wield over frontier AI releases.
Dean Ball, a former White House AI adviser who is set to join OpenAI, warned that the executive order creates a de‑facto licensing regime for cutting‑edge AI. He argued that without clear safety standards, the policy could lead to endless launch delays, potentially handing a strategic advantage to China and jeopardizing billions of dollars earmarked for AI infrastructure.
OpenAI’s blog post stressed that the company does not view the current government‑access process as a sustainable default. "We don’t believe this kind of government access process should become the long‑term default," the post read. "It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."
Nevertheless, the firm framed the restriction as a short‑term measure. It said GPT‑5.6 will return to broader availability in the coming weeks as OpenAI collaborates with the administration to craft a new executive‑order framework focused on cybersecurity and a repeatable process for future model launches.
Sol, OpenAI’s flagship model, introduces a "max" reasoning mode and an "ultra" mode that deploys coordinated sub‑agents to tackle highly complex tasks. The company claims these features push token usage higher but also deliver superior performance on benchmark tests, edging out Anthropic’s Claude Mythos 5 in coding workflows while using roughly a third of the output tokens.
To address safety concerns, OpenAI says Sol incorporates its most robust security stack yet. The model is hardened against adversarial attacks and is engineered to prioritize defensive cybersecurity work over offensive exploits. Unlike previous generations that relied on separate filtering layers, Sol’s guardrails are baked directly into its core behavior, a design choice intended to avoid the false‑positive routing that plagued Anthropic’s Fable 5.
The limited rollout also includes a tiered pricing structure: Sol costs $5 per million input tokens and $30 per million output tokens; Terra is half that price; and Luna is priced at $1 per million input tokens and $6 per million output tokens. OpenAI noted improvements to prompt caching that should make repeated queries cheaper and more predictable.
OpenAI plans to extend GPT‑5.6 access to ChatGPT users, Codex developers and API customers soon after the short‑term restriction lifts. For now, the company’s compliance with the government request underscores the growing tension between rapid AI innovation and emerging regulatory oversight.
Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.