OpenAI Delays Wider GPT-5.6 Release Following Government Request

The move comes as the U.S. government increases pressure on AI developers to place tighter controls on their most advanced systems. Earlier, after Anthropic introduced its most powerful publicly available model, Fable 5, the administration instructed the company to remove access for all foreign nationals, prompting Anthropic to withdraw the model entirely.

The situation has renewed debate over how much influence governments should have over the release of frontier AI models. Dean Ball, a former White House AI adviser who is set to join OpenAI, has argued that President Trump’s recent executive order—which asks selected AI companies to voluntarily submit their most advanced models for government review up to 30 days before launch—has effectively created an involuntary licensing system for frontier AI, resulting in increasingly restrictive oversight.

Ball also warned that the issue becomes more significant when regulators lack clearly defined safety standards. According to his view, uncertain review processes could lead to repeated launch delays that not only risk slowing the United States in its competition with China but could also threaten the billions of dollars currently being invested in AI infrastructure.

Although OpenAI has complied with the administration’s request for this release, the company made clear that it does not want this approach to become permanent.

“We don’t believe this kind of government access process should become the long-term default,” OpenAI wrote in a blog post published Friday. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”

The company described the limited preview as a “short-term step” intended to pave the way for wider availability of GPT-5.6 over the coming weeks while it works with the administration to establish a new executive order framework covering cybersecurity, along with what it described as a “repeatable process for future model releases.”

GPT-5.6 Sol specifications

OpenAI says GPT-5.6 Sol is the company’s most advanced AI model to date, delivering stronger agentic capabilities across coding, biology, and cybersecurity. The model introduces a “Max” reasoning mode alongside an “Ultra” mode that coordinates multiple subagents to tackle highly complex tasks—an approach that also significantly increases token usage.

According to OpenAI, GPT-5.6 Sol performs strongly across several industry benchmarks. The company says it slightly outperforms Anthropic’s Claude Mythos 5 on coding workflows while remaining competitive with the Mythos preview version and generating only about one-third as many output tokens.

To address concerns about the safety of increasingly capable AI systems, OpenAI said that Sol incorporates its most advanced security architecture to date. The company stated that the model has been extensively hardened against adversarial attacks and intentionally prioritises defensive cybersecurity tasks over offensive exploitation. In practice, this means the model is designed to resist jailbreak attempts while focusing on helping users understand how to defend systems rather than how to compromise them.

OpenAI also said its safety protections are integrated directly into the model itself rather than being applied through a separate filtering layer. The company appears to be avoiding the problems Anthropic experienced with Fable 5. During the brief period when that model was publicly available, requests involving higher-risk subjects such as cybersecurity, biology, or chemistry were automatically redirected to an older AI model whenever internal classifiers detected sensitive topics. That hidden routing system generated numerous false positives and prompted criticism from users.

While GPT-5.6 is initially being released only to a limited group of trusted partners, OpenAI said broader access through ChatGPT, Codex, and its API is expected in the near future.

The GPT-5.6 family will be available in three pricing tiers. Sol is priced at $5 per million input tokens and $30 per million output tokens. Terra costs roughly half as much, while Luna is priced at $1 per million input tokens and $6 per million output tokens. OpenAI also said it has enhanced prompt caching to make repeated requests more predictable and less expensive for customers.