You know that sinking feeling when you realize you’ve been paying for a Ferrari but only ever drive it to the grocery store? That’s exactly what’s happening with AI right now.
I’ve been running a simple experiment. For the past month, I’ve used DeepSeek v4 flash for every single guided agent workflow—in-IDE code generation, reviewing diffs, refactoring. Cost: about a dollar a day. My productivity? More than doubled.
Here’s the part that stings: I tried swapping in Haiku, Opus, Sonnet. The difference was indistinguishable. You are paying ten times more for a model that is, in practice, no better at your actual job.
This isn’t a “model A vs model B” debate. It’s a question the entire industry is avoiding: What are frontier models actually for?
The answer is uncomfortable. Frontier models are not designed to make you faster at writing code you already know how to write. They are designed to eliminate the need for you to write it at all. They are training for the day when a prompt says “deploy the app” and the AI does it—end to end, unsupervised, no human in the loop.
Meanwhile, the industry is pouring billions into making GPT-6 slightly better at writing a React component that you’ll still have to review. The real frontier isn’t “smarter models.” It’s autonomous task completion.
And here’s the twist that should keep you up at night: right now, you’re overpaying for a tool that makes you a better supervised developer. But the actual disruption is coming from models that don’t need your supervision at all. The $1-a-day model is already good enough to replace the junior developer who reviews diffs. The frontier model is being built to replace the senior developer who writes the diffs.
So what do you do? Stop chasing token counts. Start asking yourself: Which parts of my workflow can be automated without human oversight, and which models are optimized for that?
The answer might save you a lot of money—and your job.
FAQ
Q: Are you saying frontier models like GPT-4 and Claude are useless?
A: No. They are extremely capable, but for structured, supervised workflows (like in-IDE coding with human review), the marginal benefit over cheaper models is negligible. The real value of frontier models lies in unsupervised, autonomous tasks — which most developers aren't using them for yet.
Q: What should I do with this information?
A: Audit your AI spend. If you're using frontier models for guided agent workflows, switch to a cheaper alternative and see if you notice a difference. Then invest the savings into experimenting with autonomous agent setups — that's where the job displacement risk and ROI are both headed.
Q: Isn't this just a hot take by a single developer?
A: The experiment is anecdotal, but the logic is robust. The gap between cheap and expensive models is narrowing for common tasks. The industry's real pivot is toward unsupervised agents — and history shows that when the cost of a capability drops 10x while performance stays the same, the market flips fast.