Summary
On June 26, 2026 OpenAI previewed its GPT-5.6 model family - Sol (frontier reasoning and long-horizon agentic work), Terra (balanced, ~2x cheaper than GPT-5.5-class), and Luna (fastest, cheapest) - initially to roughly 20 organizations, with broader availability coordinated with the US government and Sol set to run on Cerebras at up to 750 tokens/sec in July.
What changed
OpenAI previewed GPT-5.6 Sol, Terra, and Luna, introducing a new max reasoning effort and an ultra mode that uses subagents to accelerate complex work; access started with about 20 partner organizations under a government-coordinated phased rollout, with general availability targeted for mid-July at the earliest.
Why it matters
A staged, government-coordinated model release plus a subagent-based ultra mode signals both tightening frontier-model governance and a shift toward multi-agent inference, shaping which models coding tools and agent platforms can build on and when.
Evidence excerpt
GPT-5.6 introduces a new max reasoning effort... and also introduces ultra mode, which goes beyond a single-agent setup by using subagents to accelerate complex work.