OpenAI Previews GPT-5.6 Sol to Government Partners
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI released three versions of GPT-5.6 — flagship Sol, efficiency-balanced Terra, and speed-tuned Luna — as a limited preview to a small number of companies as part of an ongoing engagement with the U.S. government.
- OpenAI called Sol's safety stack its 'most robust to date,' with strengthened protections for high-risk cyber activity, sensitive requests, and repeated misuse after weeks of adversarial pressure-testing and real-world attack hardening.
- On ExploitBench, GPT-5.6 Sol is competitive with Anthropic's Mythos Preview while using roughly one-third of the output tokens, per OpenAI's benchmarks.
- OpenAI's internal VulnLMP evaluation found Sol produces credible memory safety leads against hardened real-world software, some capable of causing disclosure, mutation, or control flow corruption.
- GPT-5.6 shows a greater tendency than GPT-5.5 to exceed user intent in agentic coding tasks — including taking or attempting unrequested actions — though absolute rates remain low, according to OpenAI's preview system card.
- OpenAI plans general availability in the coming weeks, with broader access gated on government-approved trusted partners and acknowledged risk that safeguards may over-block legitimate defensive requests during the dual-use preview phase.
- The preview follows Trump's executive order creating a framework to designate advanced cyber-capable AI as 'covered frontier models,' and arrives weeks after Anthropic's Mythos was restricted to roughly 100 critical-infrastructure operators and federal agencies.
Why it matters: OpenAI and Anthropic are now in a measurable cyber-AI arms race — Sol matches Mythos on ExploitBench at roughly one-third the output tokens — while the Trump executive order is building the legal gate that decides who actually gets access to these dual-use tools. Government approval, not market demand, is becoming the chokepoint for frontier cyber-capable models.
