OpenAI Limits GPT-5.6 Rollout on Gov Request

Why the government is involved at all

This did not happen in isolation. The administration has been putting new pressure on AI companies to restrict their most advanced systems. The pattern showed up sharply with Anthropic: after it released its most powerful public model, Fable 5, the administration ordered the company to remove access for any foreign national, which prompted Anthropic to pull the model down entirely.

The mechanism underneath these requests is a recent Trump executive order. It asks certain AI companies to voluntarily submit their most advanced models for government review up to 30 days before release. Dean Ball, a former White House AI advisor and soon-to-be OpenAI employee, argues that this voluntary submission has hardened into something closer to a de facto involuntary licensing regime for frontier AI.

The risk Ball flags is concrete: when the government has no clearly defined safety standards, reviews can stretch into endless launch delays, which could hand an advantage to China in the AI race and put billions of dollars in AI infrastructure spending at risk.

So the fight here is not really about one model. It is about who controls the release valve for the most capable AI systems, and whether a 30-day pre-release review becomes a standing checkpoint that every frontier launch has to clear.

How GPT-5.6 Sol is built differently

OpenAI is selling Sol on capability and on safety at the same time. On capability, it says Sol has improved agentic abilities in coding, biology, and cybersecurity. Sol introduces a max reasoning effort mode and an ultra mode that uses coordinated subagents to tackle highly complex tasks, the kind of approach that also sends token usage climbing fast.

On benchmarks, OpenAI claims GPT-5.6 is slightly better at coding workflows than Anthropic's Claude Mythos 5, a model the administration also effectively banned this month. OpenAI also says Sol is competitive with the Mythos preview while using a third of the output tokens, which would translate into real cost savings on heavy workloads.

The safety design is where OpenAI is clearly trying to learn from a competitor's stumble. The company says Sol ships with its most robust security stack yet, heavily hardened against adversarial attacks and intentionally optimized to favor defensive cybersecurity work over offensive exploits. The stated goal is a model that is hard to jailbreak and that leans toward showing users how to defend against attacks rather than how to break into systems.

The Fable 5 trap OpenAI is avoiding

OpenAI also says its safety guardrails are built directly into the core model's behavior, instead of bolted on as a separate filter. That choice reads as a direct response to what tripped up Anthropic. In the brief window that Fable 5 was available, whenever its classifiers detected a high-risk topic such as cybersecurity, biology, or chemistry, the system did not simply block the prompt. It quietly rerouted the request to an older model. That invisible downrouting produced a wave of false positives and user backlash.

Pricing, and where this goes over the next 24 to 72 hours

GPT-5.6 lands with tiered pricing. Sol costs 5 dollars per million input tokens and 30 dollars per million output tokens. Terra is half that. Luna sits at 1 dollar input and 6 dollars output per million tokens. OpenAI says it also improved prompt caching so repeated prompts are cheaper and more predictable, which matters for anyone running agentic workloads at scale.

In the immediate term, expect the trusted-partner preview to dominate the conversation while the broader public, ChatGPT users, Codex users, and API customers wait. The most important thread to watch is the framework OpenAI says it is negotiating with the administration. If that produces a clear, repeatable review process with defined safety standards, the current restriction looks like a one-time hurdle. If it does not, every future frontier launch from any major lab could face the same pre-release gate.

For developers and enterprises, the practical takeaway is to plan around uncertainty. Treat GPT-5.6 access as provisional, watch for the broader rollout in the coming weeks, and assume that government review timelines, not just engineering timelines, may now shape when the most powerful models actually ship.

Source: TechCrunch