A new benchmark tests whether AI models actually look before they leap.
An arXiv preprint (arXiv:2609.17865, not yet peer-reviewed) introduces SAFE, a benchmark that hands GPT-5.5, o3, Claude Opus 4.8, and Claude Sonnet 4.6 deployment decisions with optional safety evidence sitting nearby. The catch: checking that evidence has a cost, and the researchers varied it along with the severity and stated probability of the risk. Opus checked the evidence almost by default. o3 skipped it most, only bothering once risk crossed a clear threshold, while GPT-5.5 and Sonnet sat in between.
That's a different failure mode than the one most safety testing chases. Everyone worries about what a model does once it knows about a risk; this study says some models avoid finding out in the first place, especially when checking is costly or when it's their own deployment payoff on the line. Oddly, the models' stated reasoning didn't match what actually drove their choices: they cited probability constantly even though it barely moved behavior, and rarely mentioned evidence framing even when that swung the decision.
Take it for what it is: one un-peer-reviewed study, four models, and a synthetic decision game, not proof that any of these systems cut corners in production.