OpenAI: GPT-5.6 Sol

Sol did not launch straight to the public. On 26 June 2026, OpenAI put the finished model in front of roughly 20 organisations "at the US government's request" — thirteen days before anyone else could reach it. General availability followed on 2026-07-09source, accessed 2026-08-28: OpenAI released the family into its own API that day, and GitHub began rolling the three models out in Copilot. The same VentureBeat report on the preview carries two scores the GA announcements didn't repeat: this row measured 91.91% on TerminalBench 2.1 in ultra thinking mode, and 88.76% in max modesource, accessed 2026-08-28, and 96.7% on OpenAI's internal capture-the-flag cybersecurity testingsource, accessed 2026-08-28.

This row's own record calls it the highest reasoning ceiling in the familysource, accessed 2026-08-28, and the scoreboard agrees against its two same-day siblings. The coding index reads 77.4openrouter-models, last checked 2026-09-14 here, against 76.7openrouter-models, last checked 2026-09-14 for Terra and 71.4openrouter-models, last checked 2026-09-14 for Luna; the agentic index runs the same order: 50.5openrouter-models, last checked 2026-09-14 against 43.7openrouter-models, last checked 2026-09-14 and 42.7openrouter-models, last checked 2026-09-14. Three models, one launch day, one government-picked first audience — and Sol is the one OpenAI put at the top of it.