GPT-5.6 opens to everyone as three tiers, not one dial
On 9 July 2026 OpenAI opened the whole GPT-5.6 family — Sol, Terra and Luna — for general access in ChatGPT, Codex and the API, two weeks after a limited preview held under government supervision. The number now marks the generation and the names mark fixed capability tiers that can be updated independently of one another.
Why it matters
The company stopped selling one model with an effort dial and started selling three fixed tiers, each on its own update cadence; the name GPT-5.6 no longer points at a particular model but names a generation. It is also the first frontier release in the United States to pass through a government gate and come out of it: less than two weeks separated the request to limit access on 26 June from permission for a wide release.
The general release was announced on 9 July 2026. The family has three tiers: Sol, the flagship for hard coding and long agentic work; Terra, the balanced model for everyday tasks; and Luna, the fastest and cheapest. Prices on the day of release, per million tokens: Sol at 5 dollars for input and 30 for output, Terra at 2.50 and 15, Luna at 1 and 6. Two modes shipped above the default: max gives the model more time to think through a single answer, and ultra by default spreads a hard task across four parallel agents. The API gained programmatic tool calling, where the model writes and runs small in-memory programs that coordinate tool calls themselves and discard intermediate data. The gate opened the day before the release. On 8 July it became known that all three variants would arrive on 9 July; by contemporaneous accounts permission followed additional testing by the Department of Commerce's Center for AI Standards and Innovation and meetings for which OpenAI sent technical experts to Washington. The figures OpenAI gives are to be read as a vendor's claim. The company states that Sol sets a record of 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Claude Fable 5, and 53.6 on Agents' Last Exam. In OpenAI's own table, however, Sol scores 64.6 per cent on SWE-Bench Pro against 80 for Fable 5 and 80.3 for Claude Mythos 5, losing to both on that test. TechCrunch reports Altman telling CNBC that Sol is 54 per cent more token efficient on agentic coding, without naming what it is measured against.