OpenAI limits GPT-5.6 rollout after government request, says restrictions shouldn’t be the norm

The problem compounds, Ball argues, when the government doesn’t have clearly defined safety standards, which could lead to endless launch delays that might not only give a hand to China in the AI race, but also jeopardize the billions of dollars going to AI infrastructure buildouts. 

And while OpenAI did as the administration asked this time around, the AI firm made it clear it wasn’t happy with the arrangement.

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

OpenAI called the preview a “short-term step” that will put GPT-5.6 on the path to broader availability in the coming weeks, as the company works with the administration to develop a new executive order framework on cybersecurity, as well as a “repeatable process for future model releases.”

OpenAI says GPT-5.6 Sol is its strongest model yet, with improved agentic capabilities in coding, biology, and cybersecurity. Sol introduces a “max” reasoning effort mode and an “ultra” mode that uses coordinated subagents to solve highly complex tasks (just the sort of neat trick that sends your token usage skyrocketing).

GPT-5.6 excels at several benchmarks, says OpenAI, including being slightly better at coding workflows than Anthropic’s Claude Mythos 5, which the Trump administration also effectively banned this month. OpenAI says GPT-5.6 Sol is also competitive with Mythos preview but uses a third of the output tokens. 

To assuage any fears of its powerful models being unsafe, OpenAI says Sol includes its most robust security stack yet. It is, OpenAI says, heavily hardened against adversarial attacks and intentionally optimized to favor defensive cybersecurity work over offensive exploits. In other words, it’s designed to be hard to jailbreak, while prioritizing showing users how to defend against exploits, rather than how to hack into systems. 

OpenAI also says its safety guardrails are built directly into the core model’s behavior, rather than relying on a separate filter on top of it. The firm is likely trying to avoid the trap that caught Anthropic with Fable 5. In the brief moments when Fable 5 was available, whenever the model’s classifiers detected a high-risk topic — like cybersecurity, biology, or chemistry — it wouldn’t just block the prompt; it would route the request to an older model. The whole over-cautious flow and invisible downrouting led to many false positives and user backlash. 

While the GPT-5.6 models are initially available only to a select group of partners, OpenAI plans to make them more broadly available to people using ChatGPT, Codex, and the API soon. 

GPT-5.6 comes in three sizes with tiered pricing: Sol costs $5 per million input tokens and $30 per million output tokens; Terra costs half that; and Luna costs $1 and $6, respectively. OpenAI says it has also improved prompt caching to make repeated prompts cheaper and more predictable.