The sun, the earth, and the moon: GPT-5.6 goes public, and the velvet rope comes down

For thirteen days, GPT-5.6 existed the way a superyacht exists: you knew it was out there, you'd seen pictures, and roughly twenty government-vetted organizations were allowed aboard. That ended July 9. OpenAI opened Sol, Terra, and Luna — the sun, the earth, and the moon — to every ChatGPT user and API developer at once, and in doing so turned a preview that had been generating two weeks of secondhand speculation into something you can actually run tonight.

Three bodies, three jobs

The celestial naming isn't decoration; it's the org chart. Sol is the flagship — the frontier bid, positioned against Opus-class models and, inevitably, against Anthropic's Fable 5, which returned from its own government-ordered suspension on July 1. Terra is the workhorse, priced at $2.50 per million input tokens and $15 output — squarely at the 'everyday production workload' tier where most actual business happens. Luna is the budget play at $1 and $6, aimed at the high-volume, low-stakes work that has quietly become the largest slice of AI spending.

Read the pricing ladder next to last week's releases and the market structure snaps into focus. Grok 4.5 launched at $2 and $6 — between Terra and Luna. Z.ai's GLM-5.2 undercuts everything from below. Google's Gemini 3.5 Pro, promised for June, slipped to later this month after enterprise testers flagged reasoning issues, leaving a flagship-shaped hole in Google's lineup at exactly the moment OpenAI filled its own. Every lab is now selling a menu, not a model — and the menus are converging on the same three price points.

The frontier war used to be fought over who has the smartest model. This week it's being fought over who has the smartest three.

The Cerebras wildcard

The most technically interesting claim in the launch isn't a benchmark. OpenAI says it will serve Sol on Cerebras wafer-scale hardware at up to 750 tokens per second — roughly fifteen times the throughput of conventional GPU-based inference. If that number survives contact with production traffic, it changes what a flagship model is for: reasoning-heavy work that was previously too slow to sit inside an interactive product loop becomes something you can put behind a chat box. Treat the figure as a vendor claim until independent measurements land — but note that speed, not raw capability, is where this launch chose to plant its flag.

The velvet rope was the story too

Don't skip past how this launch was staged, because it's a template you'll see again. GPT-5.6 was previewed on June 26 and then spent nearly two weeks accessible only to a small set of government-vetted partner organizations — a controlled-release pattern that would have seemed paranoid a year ago and now reads as industry prudence. The ghost at that particular feast is obvious: Anthropic's Fable 5 launched wide on June 9 and was suspended by Commerce Department order three days later after a reported jailbreak, spending most of June dark before its July 1 restoration. Every lab watched that sequence and drew the same conclusion. A staged rollout costs you two weeks of momentum; a post-launch suspension costs you a news cycle, enterprise trust, and — as this week's market-share numbers show — measurable customer migration. The vetted-partner preview is the new airbag, and OpenAI just demonstrated it at flagship scale. Expect Google to do something similar when Gemini 3.5 Pro finally lands.

What to actually watch

Three things will tell you whether this launch mattered by August. First, independent benchmarks on Sol against Fable 5 and Opus 4.8 — the partner-wall period means almost all public numbers so far are OpenAI's own. Second, whether Terra's $2.50/$15 pricing forces a reprice from Anthropic and Google, the way Grok 4.5's pricing visibly pressured the mid-tier. Third, the Cerebras serving numbers under real load. The launch-day story is access; the real story, as always, arrives two weeks later in other people's measurements.