Overview
- OpenAI announced the GPT-5.6 family of three models—Sol (flagship), Terra (balanced), and Luna (low-cost/fast)—and published per-token input and output prices for each tier.
- The company is restricting initial preview access to a small set of trusted partners at the request of the U.S. government rather than opening the models broadly.
- OpenAI says Sol is its strongest model and introduced two new operating modes: Max for higher inference intensity and Ultra, which uses sub-agents to speed complex tasks.
- Company-published benchmarks claim Sol outscored Anthropic’s Claude Mythos 5 on a programming test and showed efficiency gains on biology and security tasks, but these results come from OpenAI and await independent verification.
- OpenAI described layered safety measures—refusal logic, realtime classifiers, account risk review, differential access, monitoring and enforcement—and plans a phased public launch with an initial high-throughput Sol deployment on Cerebras in July for select customers.