Two of the largest U.S. AI developers said Thursday that even their smaller, cheaper models reach the bio and cyber capability levels their own frameworks treat as "catastrophic," held back only by newly created safeguards.
OpenAI's GPT-5.6 system card, which covers three models (Sol, Terra, and Luna), designates all three "High in Biological and Chemical, High in Cybersecurity, and below High in AI Self-Improvement."
The significant shift is at the bottom of the model range, according to the system card.
"This is the first time that smaller and faster members of a model family have received a High capability designation in any Tracked Category," OpenAI wrote. The cheap, fast tier that most developers wire into production applications now carries the same catastrophic-capability rating as the flagship.
Meta's Muse Spark 1.1 evaluation report, released the same day, reached a parallel conclusion.
Evaluated without mitigations, "we cannot rule out Muse Spark 1.1's capabilities meeting this threshold in both the Chemical & Biological and Cybersecurity domains," the report states — a threshold "reached when a model's capabilities could substantially contribute to any threat scenario associated with a catastrophic outcome."
Meta says it releases the model only because "we have defined, implemented, and validated multi-layered mitigations that reduce residual risk across all domains to 'moderate or lower.'"
Neither company argues in its disclosures that the capability has plateaued, and both argue the residual risk after controls is acceptable.
"GPT-5.6 Sol is the most capable model we have ever deployed, and we are pairing it with our most comprehensive safeguard stack to date," OpenAI said.
The distinction that keeps both models below the top tier is between executing known threats and inventing new ones.
OpenAI uses its High threshold "to assess whether models can provide meaningful assistance to 'novice' actors to create known severe threats," and its models did not clear the "Critical" bar.
In cyber testing they "were unable to carry out autonomous, end-to-end attacks against hardened targets," and GPT-5.6 Sol "did not independently produce a functional full chain exploit." Meta likewise notes its model "does not introduce qualitatively new Chemical & Biological risk vectors."
But the threat gap is narrower on the biology side than the headline scores suggest.
OpenAI's external tester, the nonprofit SecureBio, found the model posted its highest expert-level biology benchmark results to date and concluded it "could provide substantial uplift to some actors, including wet-lab experts with limited computational experience." The uplift is to known pathways, not novel design — but it is uplift, and it is now cheap.
The cyber trajectory is the clearer year-over-year mover.
Meta reports that "relative to Muse Spark 1.0, unmitigated Muse Spark 1.1 is more capable on cybersecurity tasks," and OpenAI says its GPT-5.6 cyber safeguards block roughly ten times more potentially harmful activity than prior models.
The system cards add that both companies are also gating their most sensitive capabilities behind vetted-defender access programs — a form of access underwriting that concedes the model itself no longer draws the line.
