Summary
- OpenAI has released a limited preview of its GPT-5.6 models, named Sol, Terra, and Luna, ahead of a wider release planned for the coming weeks.
- This restricted launch was initiated at the request of the U.S. government, which is working on a framework for assessing advanced AI technologies.
- The new models feature enhanced reasoning capabilities, improved cybersecurity functions, and additional protections against misuse.
On Friday, OpenAI announced the introduction of its GPT-5.6 AI model series, featuring a limited preview of models codenamed Sol, Terra, and Luna, with a broader rollout anticipated in the near future.
This announcement follows reports that the Trump administration had requested OpenAI to restrict the initial release of GPT-5.6 while officials assess the model under a developing federal framework for advanced AI systems, confirming earlier speculation.
OpenAI has labeled GPT-5.6 Sol as its flagship model, boasting enhancements in coding, biology, and cybersecurity. Terra is designed to offer similar performance to GPT-5.5 at a reduced cost, while Luna targets high-volume, cost-effective tasks. Additionally, the launch introduces new reasoning modes, "max" and "ultra," which allow Sol more time to tackle complex challenges or manage multiple subagents in demanding scenarios.
GPT-5.6 Sol preview — it's a good model: https://t.co/EaAIeGjfIz pic.twitter.com/UihzcpfR22
— Greg Brockman (@gdb) June 26, 2026
OpenAI stated, “We are commencing a limited preview of the GPT‑5.6 series: Sol, our flagship model; Terra, a balanced model for daily tasks; and Luna, a swift and economical model.” They added, “Terra offers competitive performance to GPT‑5.5 while being twice as cost-effective, and Luna delivers strong capabilities at our lowest price point.”
In their announcement, OpenAI confirmed that they had shared the details of the models with the U.S. government prior to the launch and are starting with a limited preview due to the administration's request, while both parties work on a protocol for future advanced AI releases.
“In our ongoing collaboration with the U.S. government, we presented our plans and the models’ capabilities prior to today’s launch,” the company noted. “At their request, we are initiating a limited preview for a select group of trusted partners, the identities of which have been shared with the government, before a broader release.”
This decision follows an earlier directive from the administration that required Anthropic to restrict access to its Fable 5 and Mythos 5 models, marking GPT-5.6 as the second frontier AI system this month affected by the White House's actions.
In evaluations, OpenAI reported that Sol achieved top scores on TerminalBench, a benchmark for command-line software engineering tasks, outpacing GPT-5.5, Claude Mythos 5, and Fable 5.
“GPT‑5.6 Sol also demonstrates substantial advancements in biological workflows,” OpenAI noted. “On GeneBench v1, which assesses long-term genomic and quantitative biology analyses, it outperforms GPT‑5.5 while utilizing fewer tokens.”
Regarding cybersecurity, OpenAI stated that GPT-5.6 combines enhanced cybersecurity features with broader safeguards aimed at supporting defensive security research and curbing offensive misuse. The model remains below the Cyber Critical threshold, as it can identify vulnerabilities and exploit components but cannot autonomously create a complete exploit chain during testing.
“GPT‑5.6 is designed to deny prohibited cyber assistance, even when users try to mask their intentions or jailbreak the model,” the company explained. “These safeguards set the initial limit on what the model is permitted to assist with.”
During the preview phase, GPT-5.6 Sol, Terra, and Luna will be accessible through the API and Codex to a limited number of partners before a public rollout to ChatGPT and other users. OpenAI also introduced a new naming convention for the model family, with GPT-5.6 Sol expected to debut on Cerebras in July, achieving inference speeds of up to 750 tokens per second.
Despite the mandated limited launch, OpenAI reiterated its commitment to making the models widely available to the public.
“We advocate for broad access, and we plan to make GPT‑5.6 Sol, Terra, and Luna generally available in the coming weeks,” they affirmed.