Why $15 of Claude Credits Can Do $58,000 of Engineering Work, But Your Enterprise Can't Use It
June 6, 2025
In March 2025, Claude 3.5 completed what amounted to $58,000 of freelance SWE tasks for about $15 in API credits. Real engineering work, benchmarked against OpenAI's SWE-Lancer.
The economic arbitrage is real: for less than $15 in API credits, AI can complete software tasks worth over $58,000.
Meanwhile, McDonald's just killed their IBM Watson drive-thru because it kept adding bacon to ice cream orders. The AI worked fine. The architecture was broken.
I've been thinking about this a lot lately. Was just in Idaho at a potato processing facility helping them rip out their entire ERP system - they weren't even trying to "add AI," just wanted basic vendor automation. But watching them struggle with integrations between 30-year-old systems and modern SaaS tools... it hit me. They don't need another subscription. They need AI-native software that actually understands their business.
Here's the thing nobody talks about: enterprises maintain an average of 175 SaaS subscriptions. Only 45% see regular use. That's $21 million annually burned on software that sits idle while employees spend 20-30% of their time copy-pasting between systems.
Enterprises burn millions on idle software while employees act as human duct tape between fragmented systems.
This isn't theoretical. Real-world data shows the staggering waste of the SaaS model across major government agencies.
We built the world's most expensive game of telephone and now we're surprised AI can't play.
The Steam Engine Problem
There's this story about the Gould & Curry mine in Nevada, 1879. They had these incredible Corliss steam engines - 97% thermal efficiency, absolute peak of the technology. Electric motors were garbage by comparison. 60% efficient, unreliable, weak.
Ten years later every steam mine was either electric or dead. It wasn't about efficiency. Steam required you to build your entire operation around massive central engines and elaborate belt systems. Electricity let you put power where you needed it. The architecture mattered more than the metrics.
In the steam era, the entire factory was built around a central power source. The architecture dictated the work.
Sound familiar? Today's enterprises have perfected their steam engines. That ERP your company spent $50M implementing? It's incredibly good at what it does. Those 175 SaaS tools? Each one best-in-class. But you're still running a steam-powered factory.
What Actually Breaks
The shift from steam to electricity wasn't about efficiency; it was a fundamental architectural inversion. We are living through the same shift today.
Traditional software is deterministic. Input A → Process B → Output C. Every time. AI operates probabilistically - it pursues objectives, not rules. When you force probabilistic workloads through deterministic pipelines, things break.
The ROI crisis in AI isn't about model capability. It's about architectural mismatch. High-capability models stuck in legacy architectures lead to stalled pilots.
The models work. o3 can outcode most developers. The integration doesn't. The real problem is that only 10% of companies (McKinsey) or maybe 4% (BCG) see any ROI from AI initiatives.
Model capability is growing exponentially. AI is already outperforming the vast majority of human programmers.
I saw this firsthand working with a tax prep firm in NYC last week. They had QuickBooks, three different document management systems, and a custom Access database from 2003. Their "AI strategy" was adding a chatbot to their website. What they actually needed was to throw it all out and build something that understood tax workflows from first principles.
The New Architecture
Here's what's actually happening while enterprises debug chatbots: A team I know needed a code review tool last week. Instead of evaluating vendors, they described what they wanted to Claude. Four hours later they had a custom application integrated with their workflow. Cost: $3.40.
This isn't "no-code." It's post-code. Software that doesn't exist until you need it.
Continuous software generation for every user intent.
JPMorgan figured this out. They're not spending $15.3 billion to add AI features. They're rebuilding their entire infrastructure because they understand the physics: new computational substrate requires new architecture. That's why they're moving 75% of everything to cloud.
The companies still trying to make their 175 SaaS subscriptions talk to each other through AI middleware are missing the point. Those subscriptions are about to collapse into function calls. Salesforce becomes a prompt. Jira becomes a parameter. Your entire IT stack reducible to natural language.
The Arbitrage Window
Right now, today, you can build production systems with AI for pennies that would cost hundreds of thousands in traditional development. Some of the apps I've made have been deployed in government agencies that would require teams of developers and committes to advise.
But this window won't stay open. The same VCs subsidizing your $15 experiments today will extract monopoly rents tomorrow. We're in that weird moment where the future is here but unevenly distributed.
The potato farm in Idaho? They could automate their entire supply chain with agents that actually understand logistics, not just move data between systems. That tax firm could become a platform that transforms every small business client into a real-time financial intelligence system. But they won't, because they're still thinking in terms of features instead of foundations.
Look at your own stack. Count the integration points. Calculate how much engineering time goes to making systems talk to each other. That's what's about to become obsolete. Not the work - the plumbing.