OpenAI has announced a limited preview of its GPT‑5.6 family, consisting of three variants: Sol, Terra and Luna. The company said it shared the models and its release plans with the U.S. government and is initially making the models available to roughly 20 selected organizations; a general public release is planned "in the coming weeks."
This staged release follows an executive order issued by President Donald J. Trump on June 2, 2026, which directed federal agencies to coordinate a benchmarking and capability assessment process for new AI models within 30 days — a timeline that points to July 2 as a key milestone. OpenAI wrote in its blog that, at the government’s request, it previewed plans and capabilities ahead of the launch and is starting with a limited preview for a small set of trusted partners.
The three variants: Sol, Terra and Luna
- Sol: the highest‑capability model, targeted at the hardest problems such as complex coding, long‑running agentic workflows and security research. Pricing: $5.00 per million input tokens and $30.00 per million output tokens (the same rates as GPT‑5.5). OpenAI says Sol delivers major performance gains on extended coding, cybersecurity and agentic tasks.
- Terra: a midtier that balances strong capability with efficiency, intended for large‑scale production use cases such as customer support and document analysis. Pricing: $2.50/$15 per 1M tokens.
- Luna: the lightest and lowest‑cost option, optimized for speed and everyday tasks (summarization, drafting, routine automation). Pricing: $1/$6 per 1M tokens.
OpenAI described the new naming approach as separating generation (the numeric version) from durable capability tiers (Sol/Terra/Luna), giving clearer choices across intelligence, speed and cost. Sources said the names were chosen to evoke cosmic imagery.
Benchmarks and performance
OpenAI reports measurable performance gains over prior models in complex reasoning and long‑horizon tasks. On the TerminalBench 2.1 command‑line automation benchmark, Sol achieved a record 91.91% in ultra mode and 88.76% in max mode, outpacing GPT‑5.5’s 83.4% and exceeding Claude Mythos 5’s roughly 88% result.
In agent‑style evaluations (Agent's Last Exam), Sol was the only model to pass the halfway mark for task completion, reaching 50.9% in "code mode," while Luna narrowly outperformed the previous generation’s flagship. In quantitative biology and genomics tests, Sol and Terra recorded higher accuracy than GPT‑5.5 and GPT‑5.4, with Sol doing so while using fewer tokens.
In cybersecurity evaluations measuring vulnerability research and exploitation, the new models surpassed previous ceilings: Sol achieved higher intended exploit rates as reasoning time increased and did so using a fraction of the output tokens older models required. On ExploitBench, OpenAI says Sol performs near Mythos Preview while generating about one‑third as many output tokens.
Technical changes: longer reasoning and subagents
GPT‑5.6’s principal technical shift is giving models more time and structure for hard problems during inference. OpenAI added a max reasoning setting for GPT‑5.6 Sol for problems that need extended deliberation. The new ultra mode introduces subagents that can split and accelerate complex projects instead of keeping the work inside a single‑agent flow. OpenAI’s launch evaluations suggest this approach improves performance on several agent‑style tasks.
Predictable prompt caching and a Cerebras speed tier
The GPT‑5.6 API introduces a revised prompt caching protocol to help enterprises control cost when running agentic loops. Developers can create explicit cache breakpoints with a guaranteed 30‑minute minimum cache lifetime. Initial cache writes cost 1.25× the model’s standard uncached input rate, while later cache reads receive a 90% discount. This means initial cache establishment is more expensive, and reuse within the minimum window is much cheaper.
OpenAI is also launching Sol on Cerebras hardware in July for enterprise applications where latency is critical, claiming processing speeds up to 750 tokens per second for specialized real‑time, frontier‑grade reasoning workloads.
Security, red‑teaming and operational controls
To clear the models for release, OpenAI allocated roughly 700,000 A100e GPU hours to automated red‑teaming of GPT‑5.6, focused on discovering "universal jailbreaks" — systemic attack vectors rather than single‑prompt workarounds. The company built a multi‑layered safeguard stack that operates in real time and intentionally places operational friction for enterprise security teams.
Key safety elements include:
- Model‑level refusals: GPT‑5.6 is tuned to reject banned cyber assistance and requests that mask malicious intent.
- Live misuse screening: separate cyber and biology detectors review generations while they are produced.
- Activation‑based screening: for Sol and Terra, activation classifiers watch internal model signals during inference; if risky patterns emerge, output streaming can pause while another safety check reviews content. Luna does not appear to receive the same activation‑classifier layer but is covered by other monitoring.
- Reasoning review pauses: when elevated risk is detected, generation can stop while a larger reasoning system examines the exchange and context and blocks disallowed outputs before they reach the endpoint.
OpenAI acknowledges that legitimate defensive activities (code review, vulnerability discovery, patch engineering) use the same code primitives as offensive exploits, so classifiers may trigger false positives. The system card reports 94.8% overall recall on its biology evaluation set and 81.6% overall recall on its cybersecurity evaluation set — useful quantitative signals but indicative of imperfections.
Persistent flagging can trigger automated account‑level reviews of historical conversations to assess whether an enterprise customer is engaging in malicious behavior or standard security research. OpenAI is negotiating longer‑term enterprise safety controls, including customer‑run safety overrides and privacy‑preserving detection mechanisms to limit manual review of corporate data.
Under OpenAI’s testing, Sol is optimized for defensive containment rather than autonomous offensive deployment: when run against Chromium and Firefox codebases, the model isolated bugs and exploitation primitives but did not autonomously engineer a full, functional exploit chain. All three GPT‑5.6 models crossed OpenAI’s internal “High” cyber threshold in capture‑the‑flag testing: Sol 96.7%, Terra 91.84%, and Luna 85.19%.
OpenAI frames this as meaning the models can automate parts of vulnerability research and exploit analysis, but are not yet systems that reliably run a complete advanced attack campaign without human direction under the company’s test conditions.
Business and geopolitical implications
The phased preview and coordinated rollout with the White House underline the growing entanglement between frontier AI developers and national security protocols. OpenAI has publicly criticized the government access process used for the initial rollout, saying it should not become the long‑term default because it would keep tools from users, developers, enterprises, cyber defenders and global partners who need them.
The staged release creates a new operational environment for enterprise buyers: they must navigate real‑time safety interventions, mandatory compliance parameters, and structured token‑cache systems. The timing and approach also follow the U.S. government’s export control action against Anthropic over jailbreaks in Claude Fable 5, a move that led Anthropic to cut access to that model and its cybersecurity counterpart Claude Mythos 5.
In short, GPT‑5.6 brings technical advances in longer reasoning, subagent‑based task decomposition and enterprise cost predictability, while its initial availability, security constraints and government coordination introduce fresh operational and regulatory considerations for users and companies.
