GPT-5.6 Sol is HERE — and it Changes Everything (Terra & Luna too!)

OpenAI just previewed GPT-5.6 — not one model, a family of three. Here's the 2-minute breakdown of what's new, what the benchmarks say, and when you can touch it.

The family

Same lineup logic as every current frontier family: one to impress, one to deploy, one to spam.

What's actually new

• Terminal-Bench 2.1 SOTA. Sol takes the state of the art on the agentic terminal benchmark — the one that measures whether a model can actually drive a shell, not just chat about it.
• Biology and cybersecurity gains. Notable jumps on GeneBench and cyber evals — the dual-use frontier keeps moving.
• Two new reasoning modes. max pushes single-context reasoning further; ultra spawns sub-agents — the model orchestrates copies of itself on subtasks. Multi-agent as an API parameter, not an architecture you build.
• One-parameter API. Model + reasoning effort in a single knob; less config surface than the current maze.
• Pricing — surprisingly friendly for a flagship preview, positioned to undercut rather than premium-price.

Fun subplot: Crypto Twitter briefly lost its mind over the names (Sol/Terra/Luna — someone at OpenAI knew), and Cerebras appears in the serving story — wafer-scale inference behind the speed claims.

The catch

Limited preview. Trusted partners only, via API + Codex. General availability in ChatGPT, Codex, and the API is "weeks away." Everything public so far comes from OpenAI's own post: openai.com/index/previewing-gpt-5-6-sol.

Verdict

The headline isn't the benchmark bump — it's ultra mode. Sub-agent orchestration as a first-class API feature means the "agent framework" layer everyone's been hand-rolling starts collapsing into the model itself. If that ships at the quoted pricing, it changes how agent stacks get built. Full breakdown in the video; hands-on the day access opens.