OpenAI previews GPT-5.6 Sol with guarded rollout

OpenAI has started previewing GPT-5.6 Sol, the flagship model in a new GPT-5.6 series that also includes Terra and Luna. The company says the models are available first to selected trusted partners in the API and Codex, with broader ChatGPT, Codex and API availability planned in the coming weeks. That makes this both a model announcement and a release-policy signal: OpenAI is choosing a controlled preview for its most capable system before opening access more widely.

The product structure is straightforward. Sol is positioned as the highest-performing model for difficult work, with stronger reasoning and agentic capabilities in coding, biology and cybersecurity. Terra is the balanced model for everyday work, which OpenAI says is competitive with GPT-5.5 while being roughly twice as cheap. Luna is the fast, lower-cost option for high-volume tasks. OpenAI lists API prices at $5 input and $30 output per million tokens for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna.

The performance claims are aimed at serious users, not just chatbot demos. OpenAI says Sol reaches new state-of-the-art results on Terminal-Bench 2.1, GeneBench v1, and ExploitBench, and that it is competitive with Anthropic's Mythos Preview on challenging cyber evaluations while using far fewer output tokens. The company also describes an Ultra mode for Sol that uses multi-agent scaffolding, allowing subagents to explore parallel strategies on long-running tasks.

The safety framing is equally important. OpenAI says it trained stronger refusal behavior and is using real-time classifiers, account-level signals, differentiated access controls, monitoring and enforcement for higher-risk biological and cybersecurity misuse. It also says the models underwent automated red-teaming using more than 700,000 A100-equivalent GPU hours.

For AI users, the immediate effect is limited access: the best model may exist before most people can use it. For builders and companies, the bigger message is that frontier model releases are becoming staged, monitored and sometimes shaped by government expectations. That may slow access, but it also reflects the reality that the most capable AI systems are no longer treated as ordinary software updates.