The gate came off Thursday. OpenAI made GPT-5.6 Sol, Terra, and Luna available to everyone across ChatGPT, the API, and its Codex coding tool, ending a roughly two-week preview that had been limited — under U.S. government coordination — to about 20 vetted partner organizations. The models themselves aren’t new; the access restriction was the story when the preview began. What changed today is who gets to use them: all of you.
The lineup splits three ways. Sol is the flagship built for complex reasoning and agentic work, priced at $5 per million input tokens and $30 per million output. Terra is the balanced middle tier — OpenAI positions it as matching last month’s GPT-5.5 at half the cost, $2.50 in and $15 out. Luna is the small, fast one at $1 and $6. On OpenAI’s own numbers, the top configuration, Sol Ultra, posts about 91.9% on the Terminal-Bench 2.1 agentic-coding test — the metric the lab most wants you to notice.
OpenAI didn’t ship into empty water. The same stretch brought SpaceXAI’s Grok 4.5, which claims Opus-class performance without independent benchmarks to back it, while Anthropic’s Claude line moved too — Sonnet 5 is now the default and Fable 5 was restored on credits. Three frontier labs, one week, all shipping. The competition isn’t coming. It’s here, and it’s pricing itself aggressively.
Our take: The government gate was never really the point — it was a two-week head start for a handful of insiders. The durable signal is Terra: OpenAI is selling last month’s frontier performance at half of last month’s price. That’s the pattern now. Every release quietly re-prices the tier below it toward the floor. If you build on these models, your moat was never the model — it’s the workflow, the data, and the judgment you wrap around it, the stuff that doesn’t collapse in price every eight weeks.
What to watch
- Benchmarks vs. marketing. Sol Ultra’s 91.9% is OpenAI’s own figure, and Grok 4.5’s “Opus-class” claim has no independent proof. Wait for third-party evals before you re-platform anything.
- Launch-day capacity. Throwing a gated model open to everyone at once stresses rate limits. Watch for throttling and latency across the first 48 hours.
- The price floor. Luna at $1 in, $6 out is the cheapest frontier-family tier OpenAI has shipped. The open question is how much lower the bottom keeps sliding — and who can afford to chase it.
- The gate template. A government-coordinated preview just played out in public. Whether that becomes the standard rollout for the next flagship is a bigger story than any single model.