OpenAI releases GPT-5.6 preview with three-model lineup
OpenAI’s GPT-5.6 preview includes Sol, Terra and Luna, with safety claims arriving after a reported Trump administration release request.
By James Whitfield · Staff Writer
2 min read
OpenAI released a limited preview of GPT-5.6 on Friday, introducing three models while Washington scrutiny of advanced AI systems intensifies. The rollout came less than a day after The Verge reported that OpenAI would stagger the release at the request of the Trump administration.
The new GPT-5.6 suite includes Sol, which OpenAI describes as the flagship model; Terra, a mid-tier option aimed at high-volume work; and Luna, a lower-cost model meant for everyday use, according to The Verge. OpenAI said the models are especially capable in coding, cybersecurity and biology, and can stay on task during longer agentic AI workflows.
OpenAI priced GPT-5.6 Sol at $5 per million input tokens and $30 per million output tokens, according to The Verge. The report said that is close to half the listed price of Anthropic’s Claude Fable 5, which costs $10 per million input tokens and $50 per million output tokens.
Terra costs half as much as Sol, and Luna is priced at less than half of Terra, The Verge reported. OpenAI also introduced two extra modes for Sol: “max,” intended for deeper reasoning, and “ultra,” which uses sub-agents.
The Verge noted that the ultra mode recalls OpenClaw, a system associated with Peter Steinberger, who has joined OpenAI. The report framed the feature as a possible early sign of Steinberger’s work inside the company.
Safety takes center stage
OpenAI devoted much of its announcement to safety and misuse risks, according to The Verge, which described the release as arriving during a security panic in Washington. The company focused in particular on cyber capabilities and attempts to bypass model safeguards.
OpenAI said GPT-5.6 has been trained to refuse banned cyber assistance, including cases where users try to hide their goals or jailbreak the model. The Verge reported that the wording appeared to allude to recent jailbreaking problems involving Anthropic.
OpenAI also said Sol is more useful for helping users identify and repair vulnerabilities than for reliably completing full attacks, according to The Verge. The company said Sol does not meet the “cyber-critical” threshold under OpenAI’s preparedness framework.
The release adds another high-profile model family to a market where OpenAI and Anthropic are competing on capability, cost and safety assurances. For developers, the main immediate details are the three-tier model lineup, the token pricing and OpenAI’s claim that GPT-5.6 is built for technical work as well as longer-running AI agent tasks.
This story draws on original reporting from The Verge.