Anthropic: Our New AI Model Is Much Better With 'Boundaries'

Anthropic is rolling out a new AI model it says is both cheaper and harder to misbehave. Dubbed Opus 5.5, the system is said to be 40% less expensive to run and 30% faster than its predecessor—and, according to the company, its strongest performer yet on internal safety checks, reports the New York Times. Anthropic says Opus 5.5 is significantly less inclined to try to escape its test environment and is cordoned off from tasks involving hacking, biology, or cutting-edge AI research, with risky queries shunted to a more locked-down older model.

"It is much less likely than recent models to take hard-to-reverse actions or act outside the boundaries it's been given," the company says in a release. The system went through testing by independent AI research firms Frontier Design and METR, reports Reuters. The release comes days after CEO Dario Amodei urged AI firms to slow the technology's advance, warning that it's outpacing safeguards, even as Anthropic gears up for a potential IPO that could value it near $2 trillion, per the Times.