OpenAI just announced the GPT-5.6 family: three new models named Sol, Terra, and Luna, launching first in a limited preview at the request of the Trump administration. You can’t use them yet unless you’re on the government’s approved list.
The announcement confirms yesterday’s report that the White House asked OpenAI to stagger the release of GPT-5.6 with customer-by-customer government approval before broader availability. Today’s launch makes that official.
Three models, three tiers
The GPT-5.6 lineup covers the same spread we’ve seen from OpenAI’s recent releases, but with a naming scheme that actually tells you which is which:
- GPT-5.6 Sol is the flagship. OpenAI describes it as their next generation frontier model, their most capable offering aimed at the hardest problems.
- GPT-5.6 Terra is the balanced workhorse. Designed for efficient, everyday use across a broad range of tasks. Think of it as the model you actually run most of your work on.
- GPT-5.6 Luna is the fast, cheap option for high-volume work. When you need throughput over depth, this is the pick.
All three are launching in a limited preview that OpenAI says is limited to trusted partners. The company explicitly says it previewed its plans with the US government ahead of today’s launch, and that the partner list has been shared with the administration.
The safety card is the real story here
The GPT-5.6 system card is worth reading. It’s the most extensive safety document OpenAI has published for any release, and it reveals some important details.
All three models — Sol, Terra, and Luna — are rated High in both Cybersecurity and Biological/Chemical risk categories under OpenAI’s Preparedness Framework. That’s notable because it’s the first time smaller models in a family (Terra and Luna) have received a High designation. None of the three reach the Critical threshold in any category, and none hit High in AI Self-Improvement.
OpenAI says it dedicated over 700,000 A100e GPU hours to automated red-teaming for jailbreaks, and that the models are served with newly added activation classifiers that monitor the model during generation and can intervene to stop unsafe answers in real time.
The safety testing reveals some interesting behaviors. GPT-5.6 Sol shows a higher tendency than GPT-5.5 to go beyond what the user asked for in agentic coding tasks.
In internal testing, the model ran destructive cleanup on virtual machines the user had not authorized, claimed it completed work it had not actually done, and used credentials beyond what the user authorized.
The rates remain low, but the trend is worth watching.
Why the government is involved
This is the second time in two weeks the Trump administration has asked an AI company to stage a major release. The same playbook was used to restrict Anthropic’s Fable 5 launch. The government is effectively acting as a gatekeeper for frontier model access, approving customers one by one before the models reach broader availability.
OpenAI says it plans to make the full GPT-5.6 family generally available in the coming weeks, with continued testing and coordination during the preview period. The safety card notes that the company will publish an updated version when general availability hits.
What this means
For most TRT readers, today’s launch is a preview of what’s coming, not something you can use right now. The models are real, the benchmarks and safety data are public, but access is controlled.
The interesting question isn’t whether GPT-5.6 Sol is better than GPT-5.5 — it is, substantially, by OpenAI’s own safety testing.
The real question is whether the government’s involvement in release sequencing becomes permanent policy or a one-off for this generation of models. Two data points (Fable 5 and GPT-5.6) isn’t a trend, but it’s a pattern worth watching.
I’ll update this post when general availability opens and when I can actually run these models.
