OpenAI previewed the GPT-5.6 family on June 26. The lineup is three models codenamed Sol, Terra, and Luna, and the preview is gated behind a US-government access list of roughly 20 organizations. No Free tier. No Plus tier. No Pro tier. If your organization is not on the sheet, you do not get to see the model, and nobody outside a small set of national-security-adjacent research groups can benchmark the family independently until OpenAI opens it further.
The framing OpenAI has offered is safety and alignment testing at frontier scale, which is a reasonable position for a model that is by any account the most capable thing they have built. The framing that is louder if you have been reading OpenAI’s own charter for the last decade is that this is a very specific kind of pilot customer. AGI for the benefit of humanity, previewed to twenty organizations picked by the executive branch, is a rewrite of the pitch. It might be the right rewrite. It is a rewrite either way.
Read this against what Anthropic did the same week. Claude Sonnet 5 shipped as the default model on Free and Pro, priced at $2 in and $10 out through August, and dropped into Claude Code with higher rate limits. Claude Fable 5 came back online on July 1 after the June export-control pull, with new jailbreak severity guidelines and a HackerOne bounty attached. Anthropic is running the play where the model is available and the guardrails are public. OpenAI is running the play where the model is not available and the guardrails are the guest list. Both companies say they are optimizing for safety. They are optimizing for different definitions of it.
The awkward part for OpenAI is what happens after the access list opens. When Sol, Terra, and Luna ship broadly, they will be graded on price and throughput next to a stack of Sonnet 5 sessions that have already been running in production for two months. The competitive window that a government-only preview is designed to protect is also a window in which the rest of the market gets to normalize on the model it can actually use. If GPT-5.6 emerges at Opus prices to a market that has spent the summer running agents at Sonnet prices, the guest list will look shorter in retrospect than it did on June 26.
The remaining question is whether the twenty organizations are a preview list or a product. A frontier model that only ever ships inside a national-security perimeter is a different company than the one that spent years pitching consumer subscriptions. The next month will start to answer which one this is.