POST 01 · The numbers
I Ran 378M AI Tokens in China. My Bill Was $4.35.
This is a real billing snapshot. I ran a continuous AI workload — 378.97 million tokens, 1,491 API calls — on DeepSeek, a Chinese model. Final cost: $4.35 USD (¥31.40 CNY).
For comparison, here is the same workload at the public list prices of mainstream Western models:
| Model | Same workload, estimated |
|---|---|
| DeepSeek — what I actually paid | $4.35 |
| GPT-4o mini (budget tier) | ≈ $227 |
| GPT-4o | ≈ $843 |
| Claude Sonnet | ≈ $985 |
The three Western figures are public list prices applied to the same token count — an estimate of what this workload would have cost, not an invoice.
The gap is staggering. It echoes the global divide in gasoline prices: some regions fill up cheap, others pay a steep premium.
AI tokens are the digital fuel for autonomous agents, background automation, and long-running knowledge work. Just as cheap fuel lets every truck and factory run, cheap token pricing decides which teams — and which regions — can experiment 24/7.
This is not about one model being “better”. Capability, latency, context windows, safety guardrails all matter. But cost is the decisive factor for mass adoption.
“The best model wins the demo. The cheapest model wins the world.”
I'm Lao Li, a retired engineer in China. Every phenomenon has a reason. Subscribe to Why is China — let's understand each other better.
— Lao Li (Billy Li), Zhuhai
The working notes, by email
Free. The sources I checked, the figures I could not verify, the paragraphs that did not survive. No fixed schedule.
Unsubscribe in one click. No selling. No politics.