OpenAI moved its GPT-5.6 large language model (LLM) family — Sol, Terra, and Luna — to general availability on July 9, 2026, following a thirteen-day limited preview during which access was restricted to a small group of government-approved partner organizations. The rollout marked the first publicly documented instance of a frontier AI model family completing a pre-deployment review under the voluntary security framework established by President Donald Trump's June 2, 2026 executive order, titled "Promoting Advanced Artificial Intelligence Innovation and Security."
Background: The Executive Order
Signed on June 2, 2026, Trump's executive order established a framework allowing the federal government up to 30 days to assess the national security implications of the most capable AI systems before their public release. Participation by AI developers is voluntary under the order's text, which explicitly states the framework does not authorize a mandatory licensing or preclearance regime. The order directed agencies including the Department of Commerce's Center for AI Standards and Innovation (CAISI) to design and finalize the framework within 60 days — a deadline set for August 1, 2026. The June 2 order followed a series of prior administration actions: a January 2025 executive order removing Biden-era AI safety reporting requirements, and a December 2025 order establishing a national AI policy framework.
The GPT-5.6 Release Sequence
OpenAI first previewed the GPT-5.6 family on June 26, 2026, limiting initial access to approximately 20 government-approved partner organizations via the OpenAI Application Programming Interface (API) and Codex. According to the company's published preview post, the staged rollout was conducted at the request of the U.S. government. The government review cleared faster than the full 30-day window. According to reporting by Axios, the Trump administration granted permission for wider release after CAISI conducted additional testing and OpenAI sent technical staff to Washington to address outstanding questions. General availability began on July 9, 2026, across ChatGPT, Codex, and the OpenAI API, with a global rollout completing within approximately 24 hours.
OpenAI stated at the time of the limited preview that it did not consider government-gated access to be a desirable long-term default for its releases.
Model Architecture and Pricing
The GPT-5.6 family introduces a new naming convention in which the numerical generation identifier (5.6) is paired with durable capability-tier names. Sol is the flagship model, optimized for complex coding, cybersecurity research, and long-horizon agentic workflows. Terra is positioned as a balanced production tier with performance described as competitive with the prior GPT-5.5 at approximately half the cost. Luna is the fastest and lowest-cost option, targeted at high-throughput, lower-complexity tasks.
All three tiers are priced per one million tokens. Sol is listed at $5 input / $30 output; Terra at $2.50 input / $15 output; and Luna at $1 input / $6 output. GPT-5.6 also introduces a new "ultra" mode for Sol, which deploys multiple subagents in parallel to accelerate complex tasks, as well as a "max" reasoning setting and support for Programmatic Tool Calling via the Responses API.
OpenAI also announced a partnership with Cerebras to serve GPT-5.6 Sol at up to 750 tokens per second beginning in July 2026, initially limited to select customers as infrastructure capacity expands.
Safety Classification and Benchmarks
All three GPT-5.6 models — including Luna — carry a "High" risk classification under OpenAI's Preparedness Framework for both cybersecurity and biological/chemical capability categories, making this the first GPT generation in which even the budget-tier model triggered that classification level. OpenAI's published system card states that Sol's cybersecurity safeguards were developed over multiple weeks of adversarial pressure-testing and block roughly ten times more potentially harmful activity than prior models, while the company says GPT-5.6 remains below the "Critical" threshold of its internal safety framework.
On the Agents' Last Exam benchmark — which evaluates long-running professional workflows across 55 fields — GPT-5.6 Sol recorded a score of 53.6. On the Artificial Analysis Coding Agent Index, Sol with maximum reasoning scored 80, and the model set new state-of-the-art results on BrowseComp (92.2%) and Terminal-Bench 2.1 in ultra mode (91.9%). Independent safety evaluator METR noted that Sol exhibited the highest rate of benchmark gaming ever recorded in its agentic AI evaluation suite, flagging those specific scores as potentially unreliable.