GPT-5.6 Sol debuts with subagent ultra mode at $5 per million tokens

OpenAI is rolling out GPT-5.6, a three-tier model family that puts agentic coding, biology, and cybersecurity at the center. Sol is the flagship; Terra matches roughly GPT-5.5 performance for half the cost; Luna is the fastest and cheapest option. The company is calling this a limited preview — API and Codex access goes first to select trusted partners — and says Sol does not cross the "Cyber Critical threshold" under its internal safety framework, though it's imposing tighter safeguards than usual because of the model's raw capability jump.

Sol introduces two new modes: max reasoning effort, which pushes the model to spend more compute on multi-step problems, and ultra mode, where the model spins up subagents to tackle complex tasks in parallel. That subagent architecture is the most concrete detail OpenAI has shared about how GPT-5.6 differs under the hood from the 5.5 line. Pricing for Sol is $5 per million input tokens and $30 output; Terra runs $2.50 input and $15 output; Luna is $1 input and $6 output.

The phased rollout mirrors the GPT-4 and o1 launches — OpenAI is leaning on red-team feedback and real-world partner use before opening the floodgates. A Cerebras deployment of Sol is planned for July, targeting 750 tokens per second, also under limited access at first. That speed claim would make Sol one of the fastest frontier models on the market if it holds up in production.

What happens next depends on how the early partner cohort uses ultra mode and whether the subagent pattern proves stable enough for general release. OpenAI has not said when the limited preview ends or what the Cyber Critical threshold actually measures in quantitative terms, so the timeline for wider API access remains unclear.