Related: White House Asks OpenAI to Limit the GPT 5.6 Release Over Safety Concerns
OpenAI has previewed GPT 5.6, its newest generation of models, and rather than a single system it arrives as three distinct tiers built for different kinds of work. OpenAI introduced the lineup as a limited preview of GPT 5.6 Sol, its next generation frontier model, alongside GPT 5.6 Terra, a balanced model for efficient everyday work, and GPT 5.6 Luna, a fast and affordable model for high volume tasks. The split is a clear signal of how the company now thinks about its products, offering raw capability at the top, a practical middle, and cheap speed at the bottom rather than asking one model to be everything at once.
Sol sits at the head of the family as the system meant for the hardest problems, the kind of complex coding and security research that pushes a model to its limits. OpenAI says Sol and Terra both set new high benchmark scores, while Luna lands near the previous GPT 5.5 generation on several tests despite being positioned as the quickest and least expensive option. That spread lets a developer match the model to the task, paying for frontier reasoning only when a problem demands it and dropping to a lighter model for routine work that does not.
The pricing makes the tiering concrete. Charged per million tokens, Sol costs $5 for input and $30 for output, Terra runs $2.50 and $15, and Luna comes in at $1 and $6. The roughly fivefold gap between the top and bottom of the range is the point, since the economics of running AI at scale often matter as much as peak capability, and a company processing millions of routine requests can route them to Luna while reserving Sol for the work that justifies its cost. OpenAI also introduced more predictable prompt caching with explicit cache breakpoints and a thirty minute minimum cache life, and gave Sol new inference modes it calls max and ultra for deeper single model reasoning and multi agent coordination on complex agentic tasks.
What makes this launch unusual is not the models themselves but how they are being released. Instead of a broad public rollout, OpenAI is making GPT 5.6 available first to a narrow set of around twenty organizations, and it did so only after sharing the models and its release plans with the United States government. A wider general release is planned for the coming weeks. The arrangement is the visible result of the reporting that emerged a day earlier, when it was revealed that the Trump administration had asked OpenAI to limit the launch of GPT 5.6 over safety concerns, with federal agencies approving access on a customer by customer basis during the preview.
OpenAI has paired that restraint with an emphasis on safety, saying Sol launches with its most robust safety stack to date. The company describes strengthened protections around higher risk activity, sensitive cyber requests, and repeated attempts at misuse, and says it spent multiple weeks finding weaknesses in the model and hardening it against real world attacks before release. That framing is no accident given the moment. The whole reason the government took interest is the worry that a model this capable could be turned to cyber offense, identifying and exploiting software vulnerabilities at a speed no human could match, and OpenAI is signaling that it has built defenses to match the scrutiny.
The launch also echoes a path the rest of the industry has been walking. Earlier in the year Anthropic released its frontier cyber model only to a small group of partners, arguing it was too powerful to hand out broadly, and the same logic now shapes how OpenAI is letting GPT 5.6 into the world. Where the two diverge is that Anthropic chose restraint on its own while OpenAI is doing so at the government’s request, a distinction that says a great deal about how quickly official oversight of frontier models has hardened in a matter of months.
For developers and businesses waiting on the general release, the practical takeaway is a more granular menu of options, with a frontier model for the toughest work, a balanced model for daily operations, and a cheap fast model for everything routine. For the wider industry, the more telling detail is the gate in front of it. One of the most anticipated model releases of the year is arriving quietly and selectively, vetted before it reaches the public, and how smoothly the broader rollout follows in the coming weeks will be an early test of whether this controlled style of launch becomes the new normal for the most powerful AI.
The big picture: The AI Model Race (2026): who’s winning the frontier contest
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →