OpenAI previews GPT-5.6 Sol with stricter safeguards

OpenAI presented GPT-5.6 Sol as the strongest model in a new series, alongside cheaper Terra and Luna variants. The announcement matters technically and politically. OpenAI says the preview will first be available to a selected group of trusted partners and organizations. It also described consultations with the US government about the models and their capabilities. This shows how the release of frontier models is changing. It is no longer just a product launch. Models with stronger reasoning, coding, biological and cybersecurity capabilities sit in an area where governments watch dual-use risk. OpenAI says it does not want government access to become the permanent default, but currently sees it as a path toward broader release. Technically, GPT-5.6 Sol introduces a deeper reasoning mode and an ultra mode that uses subagents for complex tasks. In coding, the model is said to reach a new high on Terminal-Bench 2.1. In biology, it is meant to be stronger than GPT-5.5 while consuming fewer tokens. In cybersecurity tasks, OpenAI says it improves the performance-cost ratio without crossing a critical safety threshold. The pricing is tiered. Sol costs 5 dollars per million input tokens and 30 dollars per million output tokens. Terra costs 2.50 and 15 dollars, while Luna costs 1 and 6 dollars. That matters for companies because agentic work can consume far more tokens than ordinary conversation. The regulatory impact is clear. OpenAI combines staged access, generation-time controls, account monitoring, risk-based access and further testing. The more powerful models become, the more their release will resemble controlled deployment of critical technology rather than a routine software launch. Terms: A token is a unit of text or data often used for model pricing. Dual use means a technology can be used for legitimate and harmful purposes. A safety threshold marks the point where model capability requires stricter rules.

OpenAI previews GPT-5.6 Sol with stricter safeguards

The full article is available to users registered at Hard Skills