Explained: OpenAI’s GPT‑5.6 series' Sol, Terra and Luna AI models

OpenAI recently unveiled the GPT-5.6 series of artificial intelligence models even as the company moves closer to counter intense competitive pressure and stringent regulatory scrutiny.

The AI company introduced three distinct capability tiers designed to balance raw computing power with affordability: the flagship Sol, the mid-tier Terra and the hyper-fast Luna.

OpenAI GPT 5.6 limited rollout

Rather than executing a wide public rollout, OpenAI is launching the series in a tightly restricted “limited preview.” The targeted launch comes at the request of the US government, following high-profile national security scrutiny where the Trump administration utilised export controls to render rival models offline but later allowed them with some conditions.OpenAI noted it is working alongside Washington ahead of an upcoming White House cyber Executive Order, though the tech giant warned that heavy government intervention should not become the industry’s permanent default.OpenAI CEO Sam Altman said:Good news first: Sol is a smart, efficient, and a significant step forward. It is the same price as GPT-5.5. Also launching in the GPT-5.6 family is Terra, with 5.5-level performance at half the price.Bad news: at the request of the US government, it is launching today in limited preview instead of the open access launch we were planning on. We are working with the government to get to general availability as fast as we can.I think it is quite reasonable to roll out models--especially as they reach significant new levels of capability--in this way. It fits with our long-held strategy of iterative deployment. But this isn't quite the process that we think is optimal.Now we will with the government to attempt to get to a transparent, reliable process for early access, and to ensure that as long as our safeguards work as intended we can release widely. We want to be a reliable, dependable partner that works with all stakeholders, and we also want to live by our mission of benefiting all of humanity. I believe the government shares most of our goals, and that they are overall doing a good job in a very difficult situation.We will work as quickly as we can to get this model in your hands and we hope you will love it.

Understanding the lineup: Sol, Terra and Luna

OpenAI’s new naming convention separates the models (GPT-5.6) from its durable performance tiers. The family is priced per 1 million tokens, providing a clear spectrum of operational choices.

To push capabilities further, OpenAI is debuting two new processing modes for its premium Sol model. A “max reasoning effort” mode allocates extra computing time for the AI to think deeply. Furthermore, a new “ultra mode” allows Sol to break away from the limits of a single agent by autonomously deploying and managing multiple “subagents” to accelerate massive, multi-step projects.

OpenAI GPT 5.6 Sol battles Anthropic’s Mythos

The most profound advancements in the GPT-5.6 series reside in cybersecurity.

OpenAI claims Sol completely shifts the performance-efficiency frontier for vulnerability research. On ExploitBench, Sol managed to tie the performance of Anthropic’s restricted, highly controversial Mythos Preview model. Sol achieved this while utilising only one-third of the output tokens required by its Anthropic rival, representing a massive leap forward in processing efficiency.Additionally, on ExploitGym – a collaborative security benchmark built by UC Berkeley researchers – all three models (Sol, Terra, and Luna) demonstrated linear improvements in hacking and defense capabilities as their reasoning parameters were increased.As these models display advanced autonomous capabilities, OpenAI built them with its most rigid safeguard stack to date. The company stressed that Sol is intentionally engineered to be better at finding and patching code bugs than successfully executing end-to-end cyberattacks.To maintain this boundary, OpenAI is deploying a layered safety stack:Refusal training: The model is trained to spot disguised intent and automatically refuse requests for prohibited cyber assistance, OpenAI says.Real-time classifiers: Algorithms actively scan outputs as they generate. If a violation is suspected, the text pauses mid-sentence while a secondary reasoning model reviews the context before deciding to block it.Account monitoring: Thirdly, OpenAI will track risk signals across entire user histories, helping systems distinguish malicious actors from legitimate security researchers performing dual-use defensive testing.To ensure these defenses hold up against evolving threats, OpenAI is said to have burned over 700,000 A100-equivalent GPU hours on automated “red-teaming”. AvailabilityDuring the current preview phase, access to the GPT-5.6 family remains restricted to select API developers, Codex users and government-vetted partners. OpenAI plans to expand availability to ChatGPT Plus users and broader enterprise clients in the coming weeks.