The GPT-5.6 Sol rollout has launched under an unusual constraint: OpenAI is initially making its most capable new model available only to a select group of partners, after the US government asked for early access ahead of a broader release. The company is not pretending to be happy about it.
‘We don’t believe this kind of government access process should become the long-term default,’ OpenAI wrote in a Friday blog post. ‘It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.’
OpenAI framed the restricted preview as a short-term step while it works with the Trump administration to develop a new executive order framework on cybersecurity and a ‘repeatable process for future model releases.’
GPT-5.6 Sol Rollout and the Government Review Process
The political backdrop matters here. On 2 June 2026, President Trump signed an executive order titled ‘Promoting Advanced Artificial Intelligence Innovation and Security,’ directing agencies to accelerate AI-enabled cybersecurity initiatives and design a voluntary framework for pre-release government access to frontier AI models.
In practice, the order 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 adviser who is set to join OpenAI, has characterised this arrangement as creating a de facto involuntary licensing regime for frontier AI, according to TechCrunch. A Skadden analysis of the executive order notes that AI developers and critical infrastructure companies should monitor forthcoming guidance from the Cybersecurity and Infrastructure Security Agency related to the order.
OpenAI plans to make GPT-5.6 broadly available to ChatGPT, Codex, and API users in the coming weeks. The restricted start, in other words, is a product delay dressed up in policy language.
What Sol Actually Does, and What It Cannot
OpenAI says GPT-5.6 Sol is its strongest model to date, 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 tackle highly complex tasks, which comes at the cost of substantially higher token usage.
On coding benchmarks, OpenAI says Sol is slightly better than Anthropic’s Claude Mythos 5, a model the Trump administration effectively banned this month. Sol is also competitive with Mythos preview but uses a third of the output tokens.
The ExploitGym benchmark used to evaluate cybersecurity capabilities was created by UC Berkeley researchers in collaboration with OpenAI and other frontier labs, according to the GPT-5.6 Preview System Card. That same document is candid about limits: Sol and Terra can identify vulnerabilities and assemble pieces of exploits, but in cybersecurity testing they were unable to carry out autonomous, end-to-end attacks against hardened targets. Neither model reaches OpenAI’s highest internal risk classification.
The system card also discloses that internal experiments found increases in certain misaligned behaviours for Sol relative to GPT-5.5, mostly driven by the model’s greater persistence. OpenAI attributes this to Sol being more dogged in pursuing goals, a trait that cuts both ways.
On the defensive side, Sol and Terra are served with newly added activation classifiers that watch the model during generation and can intervene to block unsafe answers in real time if outputs cross safety boundaries. OpenAI says Sol’s safety guardrails are built into the core model’s behaviour rather than layered on top as a separate filter.
That architectural choice appears to be a direct response to the problem that tripped up Anthropic’s Fable 5. Whenever Fable 5’s classifiers detected a high-risk topic, the model routed requests to an older system entirely, producing a wave of false positives and user complaints before the model was pulled. OpenAI is betting that integrated guardrails will hold up under pressure where bolt-on filters did not.
Three Models, Three Price Points
GPT-5.6 comes in three sizes. Sol is priced at $5 per million input tokens and $30 per million output tokens. Terra comes in at half those rates. Luna, the lightest tier, costs $1 per million input tokens and $6 per million output tokens. OpenAI says it has also improved prompt caching to make repeated prompts cheaper and more predictable.
The tiered structure gives enterprises a path to run Sol-level reasoning on sensitive tasks while routing routine workloads through Luna to manage costs. Whether that pricing holds once the model moves from restricted preview to full release is the question worth watching over the next few weeks.
