OpenAI split one model into three — that's the part worth copying

On June 26, 2026 OpenAI previewed GPT-5.6 — but not as one model. It's three: Sol (the frontier reasoning and agentic flagship), Terra (a balanced everyday model it pitches as GPT-5.5-class at roughly half the cost), and Luna (the fastest, cheapest). Pricing per million tokens lands at Sol $5 in / $30 out, Terra $2.50 / $15, Luna $1 / $6. It's a limited preview through the API and Codex for trusted partners, with general availability "in the coming weeks." OpenAI says all three clear its internal "High" cyber threshold, with Sol saturating that eval at 96.7%.

Photo by BoliviaInteligente on Unsplash.

What I take from it

The named tiers are the story. Sol/Terra/Luna isn't branding — it's OpenAI shipping the routing decision for you, and putting a price tag on each rung so the trade-off is explicit.

This is exactly the shape I want my own systems to have:

• One task, one tier — chosen by the task, not the habit. The reflex is to send everything to the flagship. But most requests in a real product are extraction, classification, or short rewrites — Luna-shaped work. Reserve the $30-output model for the requests that genuinely need it.
• The 5x spread between top and bottom is the budget lever. Sol's output costs five times Luna's. If your traffic is 80% routine, routing it correctly is a bigger cost win than any amount of prompt optimization.
• "Coming weeks" means don't build a launch on it yet. A limited preview to trusted partners is not an SLA. I'd prototype against it and keep a generally-available model behind the same interface.

I don't need OpenAI's three names. I need three lanes in my own router, an eval that decides which lane a request belongs in, and the discipline not to default everything to the expensive one. The vendor just made the lanes visible. The routing is still my job.

Sources: Previewing GPT-5.6 Sol (OpenAI), A preview of GPT-5.6 Sol, Terra and Luna (OpenAI Help Center).