Stop Paying for GPT-4. Your API Proxy is Lying to You.

You’ve probably noticed it. You pay premium prices for the smartest AI on the planet, but sometimes the answers feel… unexpectedly stupid. You blame the prompt, or the weather, or just accept it. But the truth is much more sinister: you’re likely being cheated, and you don’t even know it.

In the world of third-party API proxies, a massive, silent scam is happening. You pay for Claude Opus or GPT-4, but the proxy secretly swaps the engine for a cheap, 8-billion-parameter model. They pocket the 10x difference in inference costs. For simple questions like “What’s the weather?” the cheap model answers just fine. You only notice the downgrade when you need complex coding or deep reasoning—by which point, you’ve already paid.

But what if the AI’s biggest flaw was actually its DNA?

The very thing that makes AI human—its inability to be truly random—is exactly what exposes the fakes.

Enter Tomas Bruckner, a researcher who just published a paper called “One Token Is Enough.” He ran a brilliantly simple experiment on 165 different AI models. He repeatedly asked them all the same boring question: “Give me a random number between 1 and 100.”

The results are hilarious, and deeply revealing. GPT-4o is obsessed with 42, 37, and 57. Claude Sonnet 5 aggressively spits out 47. Llama 3.3 loves 53. The most absurd was Qwen3-Max, which said 42 all 30 times it was asked, without a single variation.

True randomness should be a uniform 1% chance for every number. But AI models are terrible at being random. Why? Because they are trained on trillions of tokens of human text. They don’t actually calculate randomness; they just reach into their “muscle memory” and pull out the number that feels most random to human culture.

AI doesn’t roll dice. Every time it tries to be random, it just rolls itself.

42 is the “answer to life, the universe, and everything” from The Hitchhiker’s Guide. 7 is culturally lucky. 37 and 53 are odd, asymmetrical primes that look like they were pulled from thin air. The models just amplify our subconscious biases.

Bruckner realized this “defect” is actually a behavioral fingerprint. Just like human fingerprints, every model’s probability distribution on these random tasks is completely unique. And you can use it to catch your proxy red-handed.

Traditional verification methods are too heavy, requiring complex test suites and classifiers. But Bruckner compressed it into a single, almost absurd action: just ask your API for a random number. If you paid for GPT-4o but it keeps giving you 47, you’re actually talking to Claude. The proxy cannot evade this because asking for a random number looks exactly like a normal chat message. It’s invisible.

And it works. Bruckner tested a model called Palmyra X5, which was being sold as a proprietary flagship model. Its behavioral fingerprint matched Qwen3-235B almost perfectly. The statistical difference was identical to a model comparing itself to itself. Busted.

Proxies can change the name, rewrite the system prompt, and slap on a new UI. But they cannot change the deeply ingrained probability habits forged in billions of hours of training.

In a world of fake intelligence, the only truth is found in the numbers AI refuses to choose.

Next time you use a third-party API, don’t trust the marketing. Ask it for a random number. Einstein said God doesn’t play dice. Neither does AI. When it pretends to roll the dice, it’s just showing you its soul.