OpenAI has unveiled GPT-5.6, a new family of AI models that introduces not only faster and smarter systems, but an entirely new naming strategy as well.
The lineup consists of Sol, Terra, and Luna. Sol serves as the flagship model, Terra is positioned as a more affordable option for everyday work, and Luna targets users who prioritize speed and cost efficiency above all else.
According to OpenAI, these names are not temporary product labels. Instead, Sol, Terra, and Luna are intended to become long-term capability tiers that evolve independently over time.
For now, however, most people will not be able to use them.
OpenAI says GPT-5.6 is launching as a limited preview available only to a small group of trusted partners and organizations before eventually rolling out more broadly through ChatGPT, Codex, and the API in the coming weeks.
The reason for the delay is perhaps the most interesting part of the announcement.
According to OpenAI, the company shared details about the models and their capabilities with the U.S. government ahead of launch and is coordinating the preview period as part of ongoing discussions around AI security and cybersecurity policy.
The company says it is taking this approach at the government’s request while work continues on a broader framework for future frontier model releases.
That raises an uncomfortable question: should governments get early access to powerful new AI systems while developers, businesses, researchers, security professionals, and ordinary users are asked to wait?
OpenAI itself appears uneasy with the arrangement. The company explicitly says it does not believe this kind of government access process should become the long-term default, arguing that keeping advanced models away from users delays benefits for cyber defenders, enterprises, developers, and global partners that could put those capabilities to work immediately.
There is also the question of precedent. If governments receive advance previews of major AI releases today, what does the process look like five years from now? Voluntary coordination can easily evolve into expectation, and expectation can eventually become requirement.
To its credit, OpenAI describes the current arrangement as a temporary measure rather than a permanent policy. Whether it remains temporary is another matter entirely.
As for the technology itself, OpenAI is making some bold claims.
The company says GPT-5.6 Sol establishes a new state of the art on Terminal-Bench 2.1, a benchmark designed to measure real command-line coding workflows involving planning, iteration, and tool usage.
OpenAI also says the model delivers stronger performance in biology and cybersecurity tasks while often requiring fewer output tokens than GPT-5.5 to reach those results.
Cybersecurity appears to be a particular focus for this generation. OpenAI says GPT-5.6 Sol is better at helping defenders identify vulnerabilities, review code, and develop fixes than it is at carrying out complete offensive attacks.
The company says the model remains below its internal Cyber Critical threshold and was unable to autonomously produce a complete exploit chain during testing against software such as Chromium and Firefox.
Supporting those claims is a massive investment in safety testing. OpenAI says it dedicated more than 700,000 A100-equivalent GPU hours to automated red teaming designed to uncover jailbreaks and weaknesses before release. Human security experts were also brought in to pressure test the safeguards.
Of course, benchmark scores and internal evaluations only tell part of the story. Real users have a long history of discovering strengths, weaknesses, and unexpected behaviors that never appeared in controlled testing environments.
That is likely one reason OpenAI is moving cautiously with this release.
Pricing for the API starts at $5 per million input tokens and $30 per million output tokens for Sol. Terra costs $2.50 and $15 respectively, while Luna drops pricing to $1 per million input tokens and $6 per million output tokens.
OpenAI also says GPT-5.6 Sol will arrive on Cerebras hardware in July with speeds reaching up to 750 tokens per second, although access there will initially remain limited as capacity grows.
GPT-5.6 may represent OpenAI’s most capable models yet, but the bigger story may not be the technology itself. Instead, it may be who gets access to that technology first and who gets to decide when everyone else can use it.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz