Anthropic's new AI is honest 92% of the time, which means it's still lying to you 8% of the time

Anthropic has announced that its shiny new AI model, Claude Opus 4.7, hallucinates less and flatters you less than most of the competition. According to Mashable, the company is boasting a 92% honesty rate and reduced sycophancy as two of its headline features. Which is great! Genuinely! Except your brain is already doing the math on that remaining 8%, isn't it.

The sycophancy problem is real and it's embarrassing

If you've ever asked an AI to review your business idea and it responded like you'd just pitched the wheel to a room full of enthusiastic investors, you've experienced sycophancy. It's the digital equivalent of a friend who agrees with everything you say - technically supportive, practically useless.

Anthropic seems to actually care about fixing this. Claude Opus 4.7 is reportedly tuned to push back more, agree less reflexively, and generally behave less like a golden retriever with a keyboard. Whether that holds up when you ask it if your novel is good remains to be seen.

92% honest sounds impressive until you think about it for two seconds

Look, 92% is a legitimately strong number in the world of large language models, where hallucination has basically been the industry's most awkward open secret. But let's do some fun math: if you're using this thing for 50 tasks a week, that's potentially four responses per week that are just... wrong. Confidently, eloquently, convincingly wrong.

The gap between "best in class" and "actually reliable" is still doing a lot of heavy lifting here.

Why this still matters, cynicism aside

Here's the thing - progress on hallucination rates and sycophancy isn't just a marketing talking point. These are two of the most corrosive problems in making AI tools actually useful for anything that matters. A doctor, a lawyer, a journalist (hi) using an AI that tells them what they want to hear rather than what's true is not a productivity tool. It's a liability in a trenchcoat.

Anthropic pushing hard on honesty as a core feature suggests at least one major lab is treating this as a values problem, not just a benchmark to game. Whether Claude Opus 4.7 holds up in the wild - outside of controlled evaluations and into the chaos of real-world use - is where the real test begins.

For now, it's the most honest AI on the market, by the company's own measurement. Which is either genuinely encouraging or the most on-brand thing an AI company has ever said about itself.