GPT-5.6 is now the default ChatGPT — and OpenAI's pitch is that it's a security model.

OpenAI opened the GPT-5.6 family — Sol, Terra, and Luna — to general availability on July 9, with Sol now serving as ChatGPT's default model. After months of the tier being gated to roughly twenty government-vetted partner organizations, the whole world is now talking to Sol, whether it knows it or not. For most of the hundreds of millions of people who open ChatGPT, the model behind the box changed overnight, silently, with no prompt to opt in or out. That alone is worth pausing on: the default is the product for the vast majority of users, and the default just moved.

But the genuinely interesting move here isn't the capability bump. Every model launch comes with a capability bump; they've become almost rhythmic. The interesting move is the positioning. OpenAI is presenting Sol not as its smartest model but as its most advanced cybersecurity model to date — competitive with Anthropic's Mythos on exploit-generation benchmarks while, per the company, using roughly a third of the tokens to get there. In a launch cycle where every lab reflexively claims the smartest model, OpenAI deliberately chose a different flag to plant: the best at security. When a company changes the axis it competes on, that's a tell about where it thinks the ground is shifting.

What changed

Structurally, it's three models and one default. Sol is the flagship; Terra is the balanced mid-tier; Luna is the fast, cheap option — the same good/better/best laddering the whole industry has converged on, because it lets a lab serve a $20 hobbyist and a seven-figure enterprise contract from one family. The tiering matters commercially, but the number worth actually watching is the token-efficiency claim on the security benchmarks. If Sol genuinely matches a rival's exploit-finding capability at a third of the compute, that is not primarily a safety story — it's an economic one wearing safety's clothes. Efficiency at the frontier is what lets you run a capability at scale rather than as a demo, and running security tooling at scale is a large, underserved, and well-funded market.

It also arrives at a pointed moment. The U.S. government has just stood up a review regime aimed squarely at frontier models' cyber capabilities (our policy desk covers it in depth this week). Launching your flagship as a cybersecurity model, in the same month regulators begin classifying models by exactly that capability, is not a coincidence of timing. It's a company reading the room and choosing to be seen as the one bringing the shield rather than the one holding the sword.

What it actually means

Here is the tension OpenAI is navigating, and it knows it perfectly well: 'cybersecurity model' cuts both ways. The same capability that finds vulnerabilities in order to defend against them finds vulnerabilities that can be exploited. There is no version of 'exceptionally good at finding software flaws' that is purely defensive — the skill is the skill, and the intent lives in the user, not the model. This is the exact dual-use problem that got Anthropic's Fable 5 pulled offline by the Commerce Department for nineteen days last month, after researchers demonstrated it producing working exploit code. OpenAI has watched that happen to a competitor and drawn the obvious lesson.

So the defensive framing is doing real strategic work. By foregrounding 'this makes defenders stronger' and pairing the launch with a posture of government engagement, OpenAI is pre-negotiating the regulatory conversation before it becomes adversarial: we are the responsible actor, our model is a shield, we are working with you rather than around you. It's a smart play. Whether it survives contact with the government's classified benchmarking process — which will judge the capability on its own terms, not on the marketing framing — is a separate question, and one being decided behind closed doors where neither we nor you can see the criteria.

Every lab is claiming the smartest model. OpenAI claimed the safest one — and where a company chooses to compete tells you where it thinks the fight moved.

What's still unproven

The token-efficiency figures are OpenAI's own, and independent security researchers have not yet stress-tested Sol at scale. 'Competitive on exploit benchmarks' is a phrase that should make defenders and regulators equally alert until the evaluation methodology is public and reproducible — a benchmark you can't inspect is a claim, not a measurement. There's also the reliability question that shadows every model marketed for high-stakes autonomous work: a security tool that's confidently wrong is worse than no tool, because it launders a false negative as an all-clear. As always, the two weeks after a launch — when the people who didn't build the model start probing it in anger — are where the real story gets written. We'll be reading that story as it's written, not the press release.