"most capable model yet for cybersecurity," OpenAI wrote — and on June 26 the company opened a narrow window on that claim, previewing GPT-5.6 Sol to a small, government-vetted set of partners while keeping broader access tightly controlled.
A government‑gated preview at OpenAI's request
OpenAI announced the Sol preview on June 26 as the first release in its GPT-5.6 series and said it briefed US government officials beforehand. At the government's request, OpenAI limited early access to a small group of vetted partners and shared the partners' names with the government. The company framed the arrangement as short-term and described it as "the fastest route to a wider release," even as it warned the process should not become the default because it can withhold tools from defenders and enterprises that need them.
The GPT‑5.6 family: Sol, Terra, Luna — and pricing
The GPT-5.6 rollout introduces a three‑tier naming scheme. Sol is the flagship cybersecurity model; Terra is a mid-range offering; and Luna is positioned as a cheaper, faster model. OpenAI said the GPT-5.6 models are available through its API and its Codex coding tool to the selected partners during the preview. Sol is priced at $5 per million input tokens and $30 per million output tokens; Terra and Luna are billed as cheaper alternatives. OpenAI also flagged a Cerebras launch planned for July. All performance figures reported by the company come from the preview and are still being tested.
Sol's capabilities, tests, and the safeguard stack
OpenAI described Sol as stronger on long‑horizon tasks such as vulnerability research and exploitation, but not a turnkey offensive weapon. On the ExploitBench test, the company said Sol matched the performance of Mythos Preview while using about a third of the output tokens. In targeted tests against Chromium and Firefox, OpenAI reported finding bugs and "exploit building blocks" but said the model did not autonomously produce a full working exploit and therefore does not cross the "Cyber Critical" threshold in its Preparedness Framework.
To limit misuse, OpenAI said Sol includes its "toughest safeguards yet." Those measures include refusals trained into the model and real‑time classifiers designed to pause and vet risky outputs. The company detailed an extensive red‑teaming effort, reporting more than 700,000 GPU hours of automated red teaming aimed at finding universal jailbreaks, supported by human expert testing.
What this means for technologists, policymakers, and affected enterprises
- Technologists and security teams: The preview signals that models like Sol can speed long‑horizon work such as vulnerability discovery and the construction of exploit components. Teams with access will likely evaluate how Sol's token efficiency (matching Mythos Preview with fewer output tokens) fits into defensive workflows; teams without access will need to balance urgency against the limited preview window.
- Policymakers and regulators: OpenAI said it is working with the US government on a cyber executive order framework for future releases. That cooperation is already shaping who sees the models early, and regulators will be watching whether the "government‑gated" approach is temporary or becomes a template for future model rollouts.
- Affected enterprises and procurement leaders: OpenAI argued the government‑access process should not become the default because it can withhold tools that defenders and enterprises need. Yet the company also called the limited preview the quickest path to broader availability, with plans for wider release "within weeks" while it finalizes policy and safeguards with the government.
OpenAI's preview of GPT-5.6 Sol lays out a practical tradeoff: accelerate safety‑conscious distribution by negotiating access with government partners, or preserve broad early access for defenders and customers at the risk of wider exposure. The company has placed heavy engineering and operational effort behind Sol — from price tiers to 700,000+ GPU hours of red teaming — and set a concrete near‑term timeline for expansion. Whether that expansion arrives within the promised weeks and how the cyber executive order framework will formalize access are the immediate facts the industry will be tracking.
https://www.infosecurity-magazine.com/news/openai-gpt-5-6-sol-limited-preview/
