New AI 'Jev' Screens Chatbot Thoughts 500 Times Cheaper Than Rivals

When AI chatbots tackle a hard question, they often 'think out loud' first — writing a hidden trail of reasoning before giving you an answer. Safety teams read those trails to catch an AI planning something harmful, like helping build a weapon or run a scam. But checking millions of these trails with a top model is slow and expensive. A researcher on the site LessWrong tested a new model format called Jev, which skips writing sentences and instead spits out a simple confidence score — more like a smoke detector that just beeps than one that explains itself.

The test ran Jev, Anthropic's Claude Sonnet 5, and OpenAI's GPT-5.6 Luna over 2,200 examples of AI reasoning, each labeled harmless, possibly harmful, or harmful. Jev matched Sonnet on accuracy (71.5% versus 71.4%), though Luna beat both at 79.4%. Where Jev crushed the competition was speed and cost. It answered in about half a second — six times faster than Sonnet — and cost roughly 6 cents per 1,000 checks, compared with $3.17 for Sonnet. That's about 566 times cheaper.

The catch matters. When it comes to actually flagging harmful reasoning, Jev is the weakest of the three. It only wins at correctly clearing safe content. Sonnet and Luna are better at catching the categories you'd most want caught: child abuse, violence, cybersecurity threats, and hate speech. Missing real harm is a far bigger problem than a false alarm, so Sonnet remains the safer pick when accuracy counts.

Still, at that price, Jev could work as a cheap first filter — scanning everything and escalating suspicious cases to a smarter, pricier model. The author also warns that newer AI models are getting better at hiding their reasoning from monitors, so no filter is a cure-all. For now, this points toward a future where basic AI safety screening becomes nearly free.