OpenAI Launches A Restricted Preview Of GPT-5.6 For A ‘Small Group Of Trusted Companions’

OpenAI has began previewing its GPT‑5.6 collection, which will likely be out there in three variations, to a restricted variety of trusted companions. The corporate says the variant Sol is its strongest mannequin but, whereas Terra is for on a regular basis use and has an analogous efficiency to GPT‑5.5 regardless of being twice as low cost. Luna, the final variant, is the corporate’s lowest price mannequin. OpenAI plans to present them a broad launch someday within the coming weeks.

The corporate gave the US authorities a preview of GPT‑5.6 and its capabilities earlier than right this moment. It is also  by the administration’s request that it’s previewing the mannequin to a small group of trusted companions “whose participation has been shared” with the federal government. “We do not consider this type of authorities entry course of ought to develop into the long-term default,” OpenAI wrote in its announcement. It stated it is taking the “short-term step,” for now, as a result of it ensures it could actually launch its newest mannequin collection to the general public quickly. 

President Trump signed an AI cybersecurity order earlier this month, which asks corporations to current their strongest fashions for voluntary authorities overview 30 days earlier than making them publicly out there. In accordance with a current report by The New York Instances, OpenAI, Anthropic, Google, xAI and Microsoft have been giving the federal government early entry to their newest fashions even earlier than Trump signed the order. Meta was the one holdout, and the US authorities has reportedly been urging it to submit its AI fashions for analysis. 

GPT‑5.6 introduces a “max” reasoning effort, which supplies Sol extra time to cause deeply. Sol can be OpenAI’s most succesful mannequin for cybersecurity and is the best choice to assist customers discover and repair vulnerabilities. OpenAI says Sol comes with strengthened protections for high-risk actions and delicate requests. It additionally says that the corporate had spent a number of weeks discovering its weaknesses and fortifying it towards real-world assaults.

The corporate put safeguards on all of the variants, nonetheless, to ensure they maintain as much as actual adversarial stress. As well as, OpenAI educated GPT-5.6 to refuse “prohibited cyber help,” together with makes an attempt at jailbreaking the mannequin. It spent 700,000 GPU hours to search out common jailbreaks to develop measures towards them, and it pledges to implement a “rapid-response course of to breed, assess, prioritize, and remediate newly found jailbreaks.”

OpenAI’s give attention to jailbreak prevention possible stems from what occurred to Anthropic. A few weeks in the past, Anthropic suspended all entry to its Mythos 5 and Fable 5 fashions after a directive from the federal government. Whereas the corporate did not say it outright, Amazon and different corporations had reportedly notified authorities that its fashions could possibly be jailbroken and used for malicious functions. It has began lifting its entry block, although, since US authorities has simply given Anthropic permission to redeploy Mythos to a choose group of organizations. 

The corporate has priced GPT‑5.6 Sol at $5 per million enter tokens and $30 per million output, a lot lower than what Fable price when it was nonetheless out there. ($10 for enter and $50 for output for a similar quantity of tokens.) Terra prices $2.50 for enter and $15 for output, whereas Luna prices $1 for enter and $6 for output.