GPT-5.6 launches, but OpenAI is taking it slow

OpenAI announced its new GPT-5.6 family last Friday, introducing three models with solar system-inspired names: Sol, its flagship model for complex reasoning and AI agents; Terra, designed for everyday production workloads; and Luna, a lightweight option. This time around, however, the launch strategy may be getting more attention than the models themselves. Instead of giving everyone immediate access, OpenAI is starting with a small group of organizations, echoing the phased rollout Anthropic used for Mythos. It’s an unexpected change of pace for a company known for fast rollouts.

That shift caught the attention of this week’s panel on the Mixture of Experts podcast. IBM Fellow Kush Varshney noted that Sol’s safety stack appears even more layered than what Anthropic shipped with Fable. Rather than relying primarily on a single classifier to filter harmful outputs, Sol adds extra training and guardrails, then routes responses through a reasoning model that double-checks them before they go out. Varshney called this “defense in depth,” a security term for stacking independent safeguards so no single layer has to catch everything.

The panel also dug into OpenAI’s claim that Sol excels at scientific reasoning against one of the company’s own benchmarks: predicting which small molecules bind to particular protein targets, a narrow test OpenAI uses as a signal of whether the model could be misused to help design harmful biological agents. “It was 7% [accurate],” said Lauren McHugh Olende, IBM Global Program Director for AI Open Innovation, who also noted that OpenAI recently lowered its own reporting threshold on that same benchmark from 50% to 30%.

Interestingly, Sol reaches its benchmark scores while using far fewer reasoning tokens than Anthropic’s models. IBM Distinguished Engineer Chris Hay suspects that’s intentional. OpenAI, he said, may be limiting how long the model is allowed to reason, preventing extended chains of thought from wandering into riskier territory. “I wonder if OpenAI has better control of how long they reason for,” Hay said. 

The phased rollout also drew some skepticism. Olende noted that the voluntary window regulators get to review frontier models before release has shrunk from 90 days to 30, and that roughly 20 organizations reportedly have early access to Sol, a small, largely undisclosed group. Hay came at it from a different angle, arguing for openness instead: “If I really think about the ethos of open source, the best thing that you can do from a model perspective is put everything out in the wild, and then everybody will discover what the true problems are, and then you can go and fix them,” he said. “I like that as a model. I like the idea of openness.”