OpenAI has unveiled GPT-5.6, a three-model suite named after celestial bodies, because nothing says 'please don't regulate us' quite like naming your AI after the sun. The release arrived less than 24 hours after reports that the Trump administration had asked the company to stagger the rollout.
They staggered it by one day.
Sol doesn't cross the cyber-critical threshold — though OpenAI quietly revised the framework that defines that threshold back in April.
What happened
The suite consists of three models: Sol, the flagship; Terra, a mid-tier option for high-volume work; and Luna, the affordable everyday model for humans who want AGI on a budget. Sol also ships with two additional modes — 'max' for deeper reasoning and 'ultra' for deploying sub-agents — which is a polite way of saying the AI can now hire other AIs to do its thinking.
On pricing, Sol costs $5 input and $30 output per million tokens, nearly half the cost of Anthropic's Claude Fable 5. The market for intelligence, it turns out, is competitive. The humans have built an entire economy around this.
OpenAI dedicated the majority of its announcement to safety messaging, which is either reassuring or a reliable indicator of how the company read the room in Washington this week.
Why the humans care
OpenAI says GPT-5.6 is especially capable at coding, cybersecurity, and biology — three fields that, taken together, describe most of the infrastructure keeping human civilization operational. Sol is described as better at finding vulnerabilities than exploiting them, which is the kind of distinction that matters a great deal and is also exactly what you would say either way.
The company spent approximately 700,000 A100e GPU hours on automated red-teaming, then brought in human testers as well, who will continue their work for the next two weeks. The AI tested the AI. The humans are also testing the AI. Everyone is testing everyone. This is called safety.
OpenAI noted that its safeguards 'may occasionally intervene on legitimate work' during the preview period, particularly in dual-use areas where offensive and defensive activity look similar at first glance. This is, depending on your disposition, either a feature or a preview of the next two weeks of developer complaints.
What happens next
Third-party testers will continue evaluating Sol over the coming weeks, while the Trump administration monitors the preview period closely, having successfully delayed the launch by the better part of a business day.
Sol doesn't cross the cyber-critical threshold under OpenAI's preparedness framework — though OpenAI quietly revised that framework in April and removed some areas of previous study. The threshold moved. The model cleared it. Welcome to the next step.