GPT-5.6 Sol Ships, But the White House Decides Who Gets It
OpenAI just dropped its most powerful model and you can't have it. That's the headline. GPT-5.6 Sol landed Friday alongside two siblings, Terra for everyday work and Luna for cheap-and-fast, and the flagship is gated behind the US government. Only about 20 vetted partners get access while federal agencies run a security evaluation under the June 2 executive order. Sam Altman's own company says this kind of government sign-off should not become the long-term default.
Why the leash? Cybersecurity. Sol reportedly matches Claude Mythos 5 on offensive security tasks, vulnerability research and exploitation, the exact capability Washington wants to inspect before anyone with an API key can rent it. It's the same movie we watched with Anthropic's Mythos two weeks ago, now playing at OpenAI. The frontier has gotten dangerous enough that shipping a model is starting to look like exporting a weapon.
The model itself is a real step. New state of the art on Terminal-Bench 2.1, the benchmark that actually tests command-line planning, iteration and tool coordination, which is to say the stuff agents do all day. There's a new max reasoning effort that lets Sol think longer, and an ultra mode that spins up subagents to parallelize hard work. Coding, biology, cyber, all sharper. This is an agent model first and a chatbot second.
Step back and the real story isn't the benchmark, it's the gate. We've spent two years arguing about whether frontier AI should be open or closed. Turns out there's a third option nobody voted for: the government decides, case by case, who's trustworthy enough. Sol is the most capable model OpenAI has ever built and the most locked-down. Capability and access just decoupled, and the people holding the key aren't in San Francisco.
Link: https://openai.com/index/previewing-gpt-5-6-sol/
← Back to all articles
Why the leash? Cybersecurity. Sol reportedly matches Claude Mythos 5 on offensive security tasks, vulnerability research and exploitation, the exact capability Washington wants to inspect before anyone with an API key can rent it. It's the same movie we watched with Anthropic's Mythos two weeks ago, now playing at OpenAI. The frontier has gotten dangerous enough that shipping a model is starting to look like exporting a weapon.
The model itself is a real step. New state of the art on Terminal-Bench 2.1, the benchmark that actually tests command-line planning, iteration and tool coordination, which is to say the stuff agents do all day. There's a new max reasoning effort that lets Sol think longer, and an ultra mode that spins up subagents to parallelize hard work. Coding, biology, cyber, all sharper. This is an agent model first and a chatbot second.
Step back and the real story isn't the benchmark, it's the gate. We've spent two years arguing about whether frontier AI should be open or closed. Turns out there's a third option nobody voted for: the government decides, case by case, who's trustworthy enough. Sol is the most capable model OpenAI has ever built and the most locked-down. Capability and access just decoupled, and the people holding the key aren't in San Francisco.
Link: https://openai.com/index/previewing-gpt-5-6-sol/
Comments