OpenAI's New AI Models Are Here, But the Government Got First Dibs

OpenAI just dropped its latest set of AI models, the GPT-5.6 series, and they're calling the star player "Sol." You'd think a big reveal like this would be a free-for-all, but not this time. Turns out, Uncle Sam wanted a peek first. So, for now, only a select group of "trusted U.S. partners" get to play with Sol and its siblings, Terra and Luna.

Because apparently, that's where we are now: the government gets pre-release access to our future robot overlords. OpenAI says this isn't going to be a regular thing, just a temporary measure while they figure out how to navigate the whole "super-smart AI is here" landscape with federal regulators.

What's Under the Hood of Sol?

Sol is the big brain of the bunch, designed to tackle the really tough stuff. It's got a new "maximum reasoning mode" that basically tells the AI to take a deep breath and really think before answering. And then there's "Ultra mode," which sounds like something out of a sci-fi movie, where smaller AI agents team up to solve problems that would make a single AI break a sweat.

If Sol is the heavyweight, Terra is the savvy budget option, offering similar performance to the previous GPT-5.5 but at half the cost. Luna is for when you need speed and cheapness, because even AI has to balance the books. OpenAI promises these models will eventually make their way to the masses via ChatGPT, Codex, and their API, but no firm date yet.

OpenAI is particularly proud of Sol's prowess in coding, biology, and cybersecurity. It's already breaking records on coding tests and outperforming its predecessors in biological analysis. In cybersecurity, it's holding its own against rival models, doing the same work with a third of the output. Which, if you think about it, is both impressive and slightly terrifying.

But before you start picturing Sol as the ultimate cyber-weapon, OpenAI clarifies: it's better at finding and fixing vulnerabilities than launching full-scale attacks. So, more digital detective, less digital supervillain. Good to know.

Safety First (and Second, and Third)

Given the power of these new models, OpenAI has also gone all-in on safety. We're talking built-in protections, real-time misuse detection, account monitoring, and different access levels. They even spent over 700,000 hours of GPU time (let that satisfying number sink in) on automated testing, trying to "jailbreak" the models themselves before anyone else could. Think of it as a pre-release stress test by a digital demolition crew.

So, while the rest of us wait for our turn, a select few are already putting Sol through its paces. OpenAI CEO Sam Altman is reportedly eager to get these models out to everyone as soon as possible, presumably after the government finishes its initial vetting. Because nothing says "cutting-edge AI" like waiting for bureaucratic approval.