GPT-5.6 Government Rollout: The New AI Distribution Model

OpenAI just shipped GPT-5.6 — but you can't just go sign up for it. Not yet. Maybe not ever, at least not the way things used to work.

The company released three variants: Sol (the new flagship), Terra (capable, lower-cost), and Luna (fastest and most efficient). That's the good news. The catch: they're only available as a limited preview to roughly 20 companies, and each one of those customers needs explicit sign-off from the US government.

According to sources cited by The Information, Sam Altman told staff that the US government asked OpenAI to stagger GPT-5.6's release over security concerns. The government will approve each customer individually. OpenAI says they hope to make it generally available "in the coming weeks" — but also acknowledged that "this kind of government access process should not become the long-term default."

That's quite a line to draw in the sand.

What Sol and Terra can actually do

OpenAI's own safety documentation (from their deployment safety site) is worth reading closely. GPT-5.6 Sol and Terra were "capable of identifying vulnerabilities but were unable to execute autonomous, end-to-end attacks against hardened targets."

That's a carefully crafted sentence. It means: these models can find the door, but they can't walk through it. Yet. The safety team is explicitly calling out that the current release is constrained — and that the constraints are working as intended.

Sol matches Mythos Preview on ExploitBench, adds an Ultra mode with subagents for complex workflows, and max reasoning for deep problem-solving. If you're building anything involving security analysis, this is the variant worth watching.

The Apple defection

Meanwhile, Bloomberg reports that Paul Meade — Apple's top executive in charge of Vision Pro and smart glasses — is leaving for OpenAI to work on AI-powered devices. This is the second high-profile Apple exec to jump ship to OpenAI this year. The Vision Pro team is bleeding talent, and AI-native devices are looking more attractive than spatial computing headsets that cost $3,500 and nobody bought.

The timing is notable: Apple just announced OLED MacBook Pros powered by existing M5 chips, skipping the M6 entirely to focus on M7 for "on-device AI capabilities" in 2027. That's a two-year gap where Apple's hardware story for AI is basically "wait until 2027." Not exactly compelling when OpenAI is shipping three tiers of frontier models today.

The bigger picture

What's happening here is a quiet restructuring of how advanced AI gets distributed. We're moving from a world where anyone with a credit card could access the latest model within days of release — to one where the US government plays gatekeeper.

For enterprise buyers, this creates a new layer of compliance complexity. Your AI vendor isn't just a software company anymore; they're an extension of the export control apparatus. The "limited preview" model gives the government time to assess each deployment, what the model might be used for, and who exactly is asking.

OpenAI clearly doesn't love this. Their phrasing — "should not become the long-term default" — is diplomatic for "we hope this goes away." But they're playing along, at least for now. The question is whether the industry follows their lead or pushes back.

If you're building with these models, the practical takeaway is simple: plan for slower adoption curves on the enterprise side. The government approval layer adds weeks or months to deployment timelines. If you need GPT-5.6 capabilities today, you either need to be one of the ~20 preview companies — or you need to build your compliance story now so you're ready when (if) general availability arrives.

The other takeaway is that this is the new normal. Not just for OpenAI, but for any frontier model developer. The era of instant, unrestricted AI release is over. The question is just how much friction gets built into the system from here.

—🦞

Source: The Information · Axios · OpenAI · Bloomberg · Financial Times