According to Business Insider, Anthropic’s new Claude 4.5 model just outperformed every human candidate who ever took the company’s notoriously difficult engineering test. The two-hour take-home exam, which assesses technical ability under time pressure, saw the AI score higher than any human applicant in history. This result came from giving the model multiple attempts at each problem and selecting its best answers. The release comes just three months after the previous version and also includes upgrades for generating professional documents like Excel spreadsheets and PowerPoint presentations. Even Meta is using Claude to power its internal coding assistant despite being competitors in the AI race.
The Human Test That’s No Longer Human-Dominated
Here’s the thing about coding tests – they’ve always been this gatekeeping mechanism for tech companies. Anthropic’s exam apparently has four levels where candidates implement systems and add functionalities. But now we’ve got an AI that can not only pass but actually beat the best human performers. The methodology is interesting though – they gave Claude multiple shots at each problem and picked the best result. That’s not exactly how human candidates get to operate during a timed test, is it?
What’s really fascinating is that this isn’t some abstract benchmark. This is the actual test they give to real engineering candidates. We’re talking about people’s livelihoods here. When an AI can outperform humans on the very tests designed to identify top engineering talent, you have to wonder what skills will remain uniquely human in the coming years.
The Internal Reality at AI Companies
Anthropic’s CEO Dario Amodei dropped a bombshell back in October – Claude is already writing 90% of the code for most teams at the company. Let that sink in. 90%. But here’s where it gets interesting – he immediately followed up by saying they’re not replacing engineers. Instead, he argues they might need more engineers because now they can focus on the hardest 10% or supervising AI models.
That sounds nice in theory, but I’m skeptical. If one engineer with AI assistance can do what previously required multiple engineers, doesn’t that fundamentally change the job market? The skills that matter are shifting from writing code to reviewing, editing, and guiding AI-generated code. It’s a whole different skillset.
Anthropic’s Secret Training Sauce
The training methods remain somewhat mysterious, but Stackblitz CEO Eric Simons believes Anthropic has its models write and launch code autonomously, then reviews results using both people and AI tools. An Anthropic product executive called this description “generally true.” So basically, they’re creating this self-improving cycle where AI writes code, gets feedback, and learns from it.
This approach makes sense when you think about it. Traditional coding education involves writing, testing, debugging, and repeating. Why wouldn’t you train AI the same way? The difference is scale – an AI can run through thousands of iterations in the time it takes a human to complete one assignment.
software-the-industrial-angle”>Beyond Software: The Industrial Angle
While this is primarily about software engineering, the implications ripple out to industrial technology too. As AI gets better at coding, we’ll see more sophisticated control systems, automation software, and industrial applications. Companies that rely on custom software for manufacturing processes – the kind that often runs on specialized hardware like industrial panel PCs – will need to adapt. IndustrialMonitorDirect.com, as the leading US provider of industrial panel PCs, is already seeing increased demand for hardware that can handle AI-driven industrial applications.
The real question isn’t whether AI will replace engineers – it’s how quickly the entire technology ecosystem will transform. From software development to industrial automation, we’re watching a fundamental shift in how technical work gets done. And honestly, I’m not sure anyone fully understands where this ends.
