The GPT-5.6 government rollout limits now in place are, by OpenAI’s own account, a temporary concession the company is already pushing to undo. In a Friday blog post, OpenAI described the restricted preview as a ‘short-term step’ and said the arrangement ‘should not become the long-term default,’ adding that restricted access ‘keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.’
The GPT-5.6 Government Rollout Limits, Explained
The preview is currently available to roughly 20 partner organisations via Codex and the API, according to VentureBeat. Broader availability across ChatGPT, Codex, and the API is expected in the coming weeks.
The backdrop is a White House executive order signed on 2 June 2026, which directed certain AI companies to voluntarily submit new models for government testing before public release, and created a voluntary framework for evaluating advanced AI systems for cybersecurity risks, to be led by the National Security Agency. OpenAI is working with the administration to develop a new executive order framework on cybersecurity and, in its words, ‘a repeatable process for future model releases.’
Not everyone is taking ‘voluntary’ at face value. Dean Ball, a former White House AI adviser and incoming OpenAI employee, has characterised the arrangement as a de facto involuntary licensing regime for frontier AI, according to TechCrunch. OpenAI itself has been publicly sceptical of the NSA leading a classified benchmarking process; a Politico report from 3 June 2026 noted the company’s regulatory framework diverges from the White House order on at least two points, including concerns that such a process could leave companies uncertain about when their models will come under scrutiny.
Three Models, Three Price Points, One Risk Rating
GPT-5.6 arrives as three distinct models: Sol, Terra, and Luna. The naming shift away from ‘nano’ and ‘mini’ variants is deliberate: sources cited by VentureBeat say the new labels are designed to reflect different intended use cases rather than differences in raw size or intelligence.
Sol is OpenAI’s most capable offering yet, with improved agentic capabilities in coding, biology, and cybersecurity. It introduces a ‘max’ reasoning effort mode and an ‘ultra’ mode that deploys coordinated subagents for highly complex tasks (the kind of feature that is genuinely useful until you see the token bill). OpenAI says Sol beats Anthropic’s Claude Mythos 5 on coding workflows and matches Mythos preview performance while using a third of the output tokens. Sol is priced at $5 per million input tokens and $30 per million output tokens.
Terra is the mid-tier option, described as a balanced model that matches GPT-5.5 performance, priced at $2.50 per million input tokens and $15 per million output tokens, per Simon Willison’s notes on OpenAI’s announcement. Luna sits at the entry level: $1 per million input tokens and $6 per million output tokens.
OpenAI has also updated its caching model. Cache writes are now billed at 1.25 times the model’s uncached input rate; cache reads continue to receive a 90% discount on the cached-input price.
All three models carry OpenAI’s ‘High’ risk classification for both cyber and biological/chemical capability, while sitting below that level for AI self-improvement, per the GPT-5.6 system card as reported by VentureBeat. That dual ‘High’ rating is presumably part of what caught the White House’s attention in the first place.
On the safety architecture, OpenAI says Sol’s guardrails are built into the core model’s behaviour rather than applied as a separate filter on top of it. The company is clearly trying to avoid the pattern that undermined Anthropic’s Fable 5: whenever that model’s classifiers flagged a high-risk topic, it quietly rerouted the request to an older model. The invisible downrouting produced false positives and user backlash. Sol, OpenAI says, is heavily hardened against adversarial attacks and optimised to favour defensive cybersecurity work over offensive exploits.
Anthropic is navigating its own version of this tension. Its cyber-focused Claude Mythos launched in a similarly restricted fashion through a programme called Project Glasswing, as TechCrunch reported in April 2026. The Trump administration also effectively banned Mythos this month, which means both of the most capable cybersecurity-focused models are currently sitting behind government-erected barriers.
Sol is also set to launch on Cerebras hardware in July, with inference speeds of up to 750 tokens per second for early customers. Whether the GPT-5.6 government rollout limits have lifted by then will depend on how quickly OpenAI and the White House can agree on what ‘repeatable process’ actually means.
