OpenAI has launched limited access to GPT-5.6 Sol, Terra, and Luna via API and Codex for a select group of trusted partners. The company initiated this preview at the request of U.S. authorities.
https://twitter.com/OpenAI/status/2070555272230384038
OpenAI plans to make GPT-5.6 Sol, Terra, and Luna publicly available in the coming weeks. Before the launch, OpenAI presented the models' plans and capabilities to the U.S. government. At the request of U.S. authorities, the initial phase is being conducted as a limited preview, and information about the partners has been shared with the government. OpenAI stated that they do not view this arrangement as a long-term norm.
Sol is described as the flagship model, Terra as the "workhorse" for everyday tasks, and Luna as a fast and cost-effective option. According to OpenAI, Terra delivers performance comparable to GPT-5.5 at half the cost. Luna is touted as the most affordable model in the lineup.
OpenAI also claimed that GPT-5.6 Sol is currently the company's most powerful model. The series includes a max reasoning mode, which allows Sol more time for in-depth task processing, and an ultra mode that employs sub-agents to expedite complex work.
According to OpenAI, Sol set a new record in Terminal-Bench 2.1 for command-line tasks, outperformed GPT-5.5 on GeneBench v1 with fewer tokens, and was competitive with Mythos Preview on ExploitBench, using about one-third of the output tokens.
Source: OpenAI.In ExploitGym tests, the company claims that all three models improved results in cybersecurity tasks as reasoning depth increased.
OpenAI stated that GPT-5.6 Sol, Terra, and Luna have the strongest security stack in the lineup. The company estimates that GPT-5.6 Sol does not cross the Cyber Critical threshold within the Preparedness Framework. In tests with Chromium and Firefox, the model identified bugs and exploitation primitives but could not autonomously create a fully functional exploit under tested conditions.
To verify security, OpenAI utilized over 700,000 GPU hours equivalent to A100 for automated red teaming aimed at finding universal jailbreaks. During the preview phase, multi-layered measures are in place, including model-level restrictions, real-time checks, account-level signals, monitoring, and enforcement actions. The company warned that some requests may be blocked or processed more slowly due to additional verification.
OpenAI also explained the new naming scheme: the number indicates the model generation, while Sol, Terra, and Luna represent capability levels. The pricing for 1 million tokens is $5 for Sol at input and $30 at output, $2.50 and $15 for Terra, and $1 and $6 for Luna.
In July, OpenAI plans to launch GPT-5.6 Sol on Cerebras at speeds of up to 750 tokens per second. Access during this phase will also be limited to select clients as the company expands its capabilities.
As a reminder, in early June, OpenAI confidentially filed an S-1 form with the U.S. Securities and Exchange Commission for a potential initial public offering.