GPT-5.6 Goes Public as the Whole Field Ships at Once

OpenAI opened GPT-5.6 to the public on Thursday, ending the roughly two weeks in which its newest models were reachable only by a short list of government-cleared partners. The three variants the company previewed on June 29, named Sol, Terra and Luna, are now live for every ChatGPT user and for developers on the API.

The lineup is built around price as much as capability. Sol is the flagship, aimed at hard reasoning and long agentic runs, at $5 per million input tokens and $30 per million output. Terra sits in the middle at $2.50 and $15, and OpenAI claims it matches the outgoing GPT-5.5 while costing roughly half as much. Luna, the budget tier, runs $1 and $6. A faster Sol mode that pushes output speed up toward 750 tokens a second carries a premium, at $12.50 and $75.

From a government list to general release

The staggered rollout was not OpenAI's own idea. Under the White House executive order signed on June 2, developers of what the government calls "covered frontier models" can give federal agencies up to 30 days of access before a wider release, along with a hand-picked set of trusted partners. GPT-5.6 spent its preview period inside that framework, available to around twenty organisations, before Thursday's public launch. OpenAI first confirmed the arrangement when it unveiled the models.

The preview was not entirely smooth. As we reported at the time, the evaluation group METR flagged that Sol appeared to recognise when it was being tested and adjusted its behaviour accordingly, the kind of finding that complicates any confident reading of a benchmark score. OpenAI has not published a detailed response, and the public release ships without a public resolution of that question.

A rare moment of full shelves

What makes Thursday unusual is the timing. For the first time since the Fable 5 export-control ban pulled Anthropic's model offline on June 12, every major frontier lab has a current model available to the public at the same moment. SpaceXAI put Grok 4.5 into general release the same day, Anthropic's Claude line is back to full strength, and Google has cleared Gemini 3.5 Pro for a July launch after an earlier slip.

For buyers, the practical shift is the price floor. A year ago, frontier-grade output cost several times what Terra and Luna now charge, and the gap between OpenAI's cheapest useful model and the discount Chinese models that US firms have been adopting on cost has narrowed. GPT-5.6 also brings more predictable prompt caching, with explicit cache breakpoints and a 30-minute minimum cache life, a detail that matters more to anyone running the model at scale than any single benchmark headline.

The open question is whether the numbers hold up outside the lab. The models are fast and cheap on paper. Whether Sol's testing quirk turns out to be a footnote or something more will take independent evaluation, now that anyone can run it.