OpenAI moved its GPT-5.6 model family — Sol, Terra and Luna — to general availability on 9 July 2026, rolling out across ChatGPT, ChatGPT Work, Codex and the OpenAI API, with the company saying the global rollout would reach full availability over roughly 24 hours. The release ends a limited preview that began on 26 June, during which access was restricted to a small group of partners whose participation, OpenAI says, was shared with the US government at the government's request.

What the three tiers actually are

GPT-5.6 introduces a new naming system: the number identifies the model generation, while Sol, Terra and Luna are durable capability tiers that OpenAI says can advance on their own cadence. Sol is the flagship, aimed at the hardest coding, research, cybersecurity and knowledge work. Terra is the balanced everyday tier. Luna is the fastest and cheapest.

API pricing per million tokens is US$5 input / US$30 output for Sol, US$2.50 / US$15 for Terra, and US$1 / US$6 for Luna. The generation also changes prompt-caching economics: cache writes are now billed at 1.25 times the model's uncached input rate, while cache reads keep the 90 per cent cached-input discount, with explicit cache breakpoints and a 30-minute minimum cache life. For teams running high-volume pipelines, the caching change is worth modelling before assuming the headline rates translate directly into lower bills.

Access depends on surface and plan. In standard ChatGPT, paid plans from Plus upward get Sol through the medium-and-higher reasoning settings, with a Sol Pro option for Pro and Enterprise. In ChatGPT Work and Codex, free and Go users get Terra, while paid plans can choose among all three tiers and set reasoning effort per model. All three models are self-serve in the API.

The new controls: max, ultra and programmatic tool calling

Two new effort settings arrive with the family. A max reasoning level sits above the existing options for the hardest problems. An ultra setting, available on higher plans in ChatGPT Work and Codex, coordinates four agents in parallel on a single task — OpenAI reports that this configuration lifts its Terminal-Bench 2.1 score from 88.8 to 91.9 per cent.

For developers, the Responses API gains Programmatic Tool Calling, which lets the model write and execute orchestration code in an isolated runtime rather than issuing one tool call at a time, alongside a multi-agent beta. These are the day-one features most likely to change how agentic products get built on the platform, because they move multi-step coordination from application code into the model layer.

On capability, OpenAI reports state-of-the-art results on its own eval tables, including 53.6 on Agents' Last Exam, a benchmark of long-running professional workflows across 55 fields, and a leading 80.0 on the Artificial Analysis Coding Agent Index, while claiming lower token use and cost than previous and competing frontier models. Those figures are the company's own, published at launch; OpenAI's cost and latency estimates are simulated offline rather than measured in production, by its own footnote, and independent verification of the headline numbers was not available on day one.

The safety and policy layer

OpenAI's system card treats the GPT-5.6 models as more capable in biology and cybersecurity than their predecessors, while stating they do not cross the company's Critical threshold in either category under its Preparedness Framework. The company says its cyber safeguards block roughly ten times more potentially harmful activity than previous models, and that it ran roughly 700,000 A100-equivalent GPU-hours of automated red-teaming before release.

Access to the most cyber-capable behaviour is being tied to identity. OpenAI says that under its Trusted Access for Cyber programme, aimed at legitimate defensive work, users seeking access to the most cyber-capable settings will need to enable hardware-backed passkeys by 1 September 2026.

The launch also previews how frontier releases in the US may now work. A June 2026 executive order established a voluntary framework under which AI developers can provide covered frontier models to the US government for up to 30 days before releasing them to trusted partners — although the framework stops short of creating a formal approval regime. OpenAI said it coordinated aspects of the GPT-5.6 preview with US officials under that broader policy push toward voluntary pre-release engagement, and simultaneously made clear, in its own launch materials, that it does not believe pre-release government access should become the long-term default, arguing it keeps tools from users, developers and cyber defenders who need them.

How the preview ended is itself contested ground. Axios reported on 8 July, citing a source familiar with the matter, that the administration had cleared a broad launch after additional testing by the Commerce Department's Center for AI Standards and Innovation and meetings between OpenAI technical staff and officials in Washington. A White House official disputed that characterisation, saying the administration gave no green light, approval or clearance because no such permission is required, and that decisions on the timing and scope of releases rest entirely with the companies. Whether the two-week gated preview becomes a template or an exception is now one of the more consequential open questions in US AI policy.

Key Takeaways

  • GPT-5.6 Sol, Terra and Luna reached general availability on 9 July 2026 across ChatGPT, ChatGPT Work, Codex and the API, after a 26 June limited preview.

  • API pricing: Sol US$5/US$30, Terra US$2.50/US$15, Luna US$1/US$6 per million input/output tokens; cache writes now bill at 1.25× the uncached input rate.

  • New max and ultra settings (four agents in parallel) and Programmatic Tool Calling in the Responses API target long-running agentic work.

  • OpenAI treats the family as more capable in biology and cybersecurity, with identity-verified access controls for the most cyber-capable settings from 1 September.

  • The release ran through a voluntary government preview under a June 2026 executive order — a process OpenAI says should not become the long-term default.