Quick Facts
- OpenAI launched GPT-5.6 Sol, Terra, and Luna on June 26, with access limited to roughly 20 organizations during an initial preview period.
- The Office of the National Cyber Director asked OpenAI to restrict the rollout after Sol crossed the “High” cybersecurity risk threshold under the company’s Preparedness Framework.
- Sol is priced at $5 input and $30 output per 1 million tokens, Terra at $2.50 input and $15 output, and Luna at $1 input and $6 output.
OpenAI unveiled its GPT-5.6 model family on June 26, introducing three tiers under new names: Sol, the flagship; Terra, a balanced option; and Luna, a fast and affordable model. The company plans a general release in the coming weeks, but an explicit request from two White House offices has slowed the initial rollout.
On June 25, the Office of the National Cyber Director and the Office of Science and Technology Policy asked OpenAI to limit GPT-5.6’s early access. The request follows a Trump executive order signed June 2, titled “Promoting Advanced Artificial Intelligence Innovation and Security,” which gives government cybersecurity teams up to 30 days to evaluate frontier AI models before broad release.
OpenAI had been briefing the government on GPT-5.6 for about a month, including in meetings CEO Sam Altman held with the White House in early June. The company shared the models and release plans with federal officials before going public.
A New Naming System
The GPT-5.6 launch introduces a permanent naming structure. The number identifies a model’s generation. Sol, Terra, and Luna identify capability tiers that can advance independently. OpenAI says this gives developers clearer choices across intelligence, speed, and cost without forcing a single model choice.
Sol targets the hardest problems, including complex coding, security research, and biology. Terra handles high-volume business tasks such as customer support, document analysis, and internal tools. Luna covers faster, lower-cost work like drafting, summarization, and routine automation. Terra carries competitive performance to GPT-5.5 at half the price.
Technical Specifications
Sol ships with a 1.5 million token context window, up 43% from GPT-5.5’s 1.05 million. It also adds two new reasoning modes. Max mode deepens single-pass reasoning. Ultra mode deploys subagents to split and accelerate complex projects.
On Terminal-Bench 2.1, a benchmark for command-line and coding workflows, Sol Ultra scored 91.9% and plain Sol scored 88.8%. Claude Mythos 5 scored 84.3%, and Gemini 3.1 Pro Preview scored 70.7%.
OpenAI is also launching Sol on Cerebras hardware in July, targeting speeds of up to 750 tokens per second. The company says that brings frontier-level intelligence to customers at speeds not previously available.
GPT-5.6 also introduces more predictable prompt caching, with support for explicit cache breakpoints and a 30-minute minimum cache life. Cache writes bill at 1.25x the uncached input rate. Cache reads continue to receive a 90% discount.
Safety and Cybersecurity Risk
Sol scored 96.7% on OpenAI’s internal cyberattack challenge test, crossing the “High” cybersecurity risk threshold under its Preparedness Framework. This is the first time the smaller, faster members of a model family have received a High designation in any tracked risk category. Sol, Terra, and Luna all carry High ratings in both cybersecurity and biological and chemical risk.
OpenAI says the models do not reach the Critical threshold. In testing, Sol and Terra could find vulnerabilities and exploit fragments but could not carry out autonomous, end-to-end attacks against hardened targets. The company also says safety guardrails are built into the core model behavior rather than applied as a separate filter.
What This Means for Software Companies
The tiered pricing structure gives development teams a direct path to match model cost with task complexity. Teams running high-volume support or document workflows can move to Terra and cut API costs in half compared to GPT-5.5. Luna offers further savings for simpler, speed-sensitive applications.
The government preview period sets a new precedent. Under the June 2 executive order, federal agencies are designing a voluntary framework by August 1 for frontier AI developers to engage with the government before broad releases. Software companies building on OpenAI’s API should expect similar delays around future model launches until that framework is in place.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.