OpenAI has previewed GPT-5.6 Sol, a new flagship model it describes as its "most capable model yet for cybersecurity." Access however is restricted to a small group of vetted partners at the request of the US government.
Announced on June 26, the Sol preview is the first release in OpenAI's GPT-5.6 series. The company said it briefed US government officials on the models beforehand and, at their request, allowed preview access with partners whose names it shared with the government.
The series introduces a new naming scheme with three tiers: the flagship Sol, a mid-range model called Terra and a cheaper, faster model named Luna.
OpenAI said it plans broader availability within weeks, as it works with the government on a cyber executive order framework for future releases.
A Government-Gated Preview
The company framed the limited rollout as a short-term step, not a permanent arrangement. It said such a government-access process should not become the default, arguing that it keeps the best tools from defenders and enterprises who need them. Nevertheless, OpenAI called the approach the fastest route to a wider release.
For now, the GPT-5.6 models are available through OpenAI's API and its Codex coding tool to selected partners.
Read more on OpenAI's cyber push: OpenAI Expands Daybreak to Help Defenders Patch Flaws
OpenAI called Sol its most capable model yet for cybersecurity, citing gains in long-horizon work such as vulnerability research and exploitation. On the ExploitBench test, the company said Sol matched Mythos Preview while using about a third of the output tokens.
Still, OpenAI said the model does not cross the Cyber Critical threshold in its Preparedness Framework. In tests against Chromium and Firefox, it found bugs and exploit building blocks, but did not autonomously produce a full working exploit.
A Bigger Safeguard Stack
Sol is reportedly better at helping defenders find and fix flaws than at carrying out attacks end to end, OpenAI said, and it came with the company's toughest safeguards yet. These include refusals trained into the model and real-time classifiers that can pause and vet risky outputs.
The company said it spent more than 700,000 GPU hours on automated red teaming to find universal jailbreaks, attacks that work across many prompts, backed by human expert testing.
Sol is priced at $5 per million input tokens and $30 per million output, with cheaper Terra and Luna tiers plus a Cerebras launch planned for July. All the performance figures are OpenAI's own, from a preview that it said is still being tested.
Image credit: Novikov Aleksey / Shutterstock.com