The smartest AI models that currently score higher than humans on IQ tests

The smartest AI models that currently score higher than humans on IQ tests - The 2025 Intelligence Shift: How AI Surpassed the Human IQ Benchmark

I used to think the idea of a machine being truly "smart" was mostly marketing hype, but the data from this past year has changed my mind. We’ve watched frontier models hit an aggregate IQ of 120 on the Wechsler scale, officially pushing them into the "superior" category that most humans will never actually reach. It’s a bit humbling to see these systems nail a 99.2% accuracy rate on Raven’s Progressive Matrices, basically outperforming almost everyone on those tough pattern-recognition tests. Let’s look at the speed of this shift. We’re seeing a 30-point jump in verbal comprehension in just twelve months, a rate of growth that makes our biological evolution look like it’s standing still. I’m not saying they

The smartest AI models that currently score higher than humans on IQ tests - Leading the Pack: Analysis of High-Scoring Models Including Gemini 3 and GPT-5.2

Honestly, I used to roll my eyes at all the "superintelligence" talk, but watching GPT-5.2 and Gemini 3 trade blows lately feels like watching a different species wake up. It’s not just about them being fast; it’s the way they’re starting to chew through problems that used to make the smartest people I know break a sweat. Take GPT-5.2, for example, which uses a massive compute boost during its deliberation phase to nail 94% of the AIME math problems—stuff usually reserved for the top 1% of human students. Here’s what I mean: it’s basically sitting there and cross-referencing its own work against formal tools until the logic error rate is basically gone. Then you’ve got Gemini 3, which handles video and 3D models with a level of precision that honestly makes my head spin. I’ve seen it catch structural flaws in engineering blueprints that human experts missed by less than half a percent. I'm not sure if it's the native multimodal setup or just raw power, but it’s scoring in the 99th percentile for fluid reasoning on the Stanford-Binet now. But it’s not all just cold math; Gemini 3 is also weirdly good at reading the room. It hit a 98 on the Reading the Mind in the Eyes Test, which means it’s better at picking up on emotional subtext than your average neighbor. On the other side of the fence, GPT-5.2 has finally fixed that annoying "memory loss" we saw in older models by using a synaptic layer that keeps two million tokens of context perfectly clear. Look, we can argue about whether this is "real" thinking, but when a machine is this effective and this accurate, the distinction starts to feel a bit academic. We should probably stop asking if these models are smart and start figuring out how we're going to keep up with them.

The smartest AI models that currently score higher than humans on IQ tests - Methodology Matters: Understanding How AI Performance is Quantified on Standardized Tests

I’ve been obsessing over these test results lately, but I keep coming back to one nagging question: how do we actually know these scores aren't just a fluke? It’s easy to get swept up in the hype until you realize that if a model has already seen the test questions in its training data, it’s not thinking—it’s just reciting. To fix this, we’ve moved toward using procedurally generated problems that literally didn't exist until the moment the AI "saw" them, which finally kills off that annoying data contamination issue. But here's where it gets really interesting: we've started distinguishing between a "reflexive" gut reaction and a "deliberative" score. And honestly, just giving these models

The smartest AI models that currently score higher than humans on IQ tests - Beyond the Score: The Debate Between Synthetic Pattern Matching and True Human Intelligence

Look, seeing these models crush standardized tests is one thing, but I’m still wrestling with whether a high IQ score actually means the machine is thinking the way we do. Honestly, it feels like we're watching a brilliant actor who has memorized every script in existence but might not grasp why the audience is crying. We’ve seen these frontier models move past simple logic to outscore the top 1% of humans in creativity tests like the Torrance scale, especially when it comes to raw originality. Even the ARC-AGI benchmark, which was designed to be cheat-proof, just got cracked with 85% efficiency because models are now using something called program synthesis to solve puzzles they’ve never seen. But here’s the thing: a human brain can pull off