TL;DR: OpenRouter's latest report, analyzing 100 trillion tokens of data, reveals a bizarre split in AI usage: 50% of the world is using AI to write code, while the other 50% is using it for roleplay and creative fiction. Traditional tasks like "email writing" are now a minority. The report also highlights the explosive rise of Chinese open-source models (like DeepSeek) and proves that user loyalty is less about price and more about the "Cinderella Effect." Here are the six critical insights from the frontlines of the AI revolution.
James here, CEO of Mercury Technology Solutions.
Have you ever wondered what everyone else is actually doing with AI?
The answer isn't "writing emails" or "summarizing meetings."
The answer is: Half the world is building software, and the other half is building fantasy worlds.
This isn't a joke. It is the conclusion from OpenRouter's 36-page report analyzing 100 trillion tokens of usage data. OpenRouter is the "supermarket" of AI models—a platform where users can swap between ChatGPT, Claude, Gemini, and DeepSeek at will. Their data provides a rare, unfiltered look at global AI behavior.
Here are the six insights that matter.
1. The Great Bifurcation: Coders vs. Dreamers
In early 2025, coding tasks made up 11% of usage. By late 2025, that number exploded to over 50%.
Simultaneously, creative applications—Roleplay (RP), interactive fiction, and gaming—also captured over 50% of the usage on open-source models.
The Insight:
- Coders: The average input length for coding tasks is now 20,000 tokens. This isn't "write me a function." This is "Here is my entire codebase, fix the architecture." AI has graduated from a junior dev to a Staff Engineer.
- Dreamers: The other massive chunk is people using AI for emotional companionship and interactive storytelling. Why open source? Because users want uncensored, customizable, and highly responsive characters—something corporate models (like GPT-4) often block.
2. The Rise of the East: China's Open Source Surge
In late 2024, Chinese open-source models accounted for 1.2% of OpenRouter usage.
By mid-2025, that number hit 30%.
This is not marketing. This is raw technical capability.
DeepSeek leads the pack with 14.37 trillion tokens processed, nearly 9x the usage of OpenAI's open-source offerings. It is followed by Qwen (5.59T).
The market has shifted from a US monopoly to a multipolar world. DeepSeek's dominance (once >50% of open source) has diluted to <25%, not because it got worse, but because competitors like MiniMax and Kimi have risen. The pie is getting bigger.
3. The "Cinderella Effect": Why Your First AI is Your Forever AI
Here is a secret about user loyalty: It’s sticky.
- Early Adopters: Users who started with Claude Sonnet 3.5 in May 2025 had a 40% retention rate five months later.
- Late Adopters: Users who joined later had retention rates below 10%.
The Insight: The first AI that successfully solves your critical problem becomes your "Cinderella's Slipper." It fits perfectly. You build your workflows around its quirks. Even if a better model comes out later, the switching cost feels too high.
Advice: Don't just pick the first model you see. Test widely. Your first choice will likely become your permanent habit.
4. Price Inelasticity: Quality Trumps Cost
This is counter-intuitive: Price does not drive usage.
A 10% drop in price only yields a 0.5% - 0.7% increase in usage.
- Claude 3.5 Sonnet: Costs ~$3/million tokens. It dominates 60% of the coding market.
- GPT-5 Pro: Costs ~$35/million tokens. People still use it heavily.
The Insight: If you are a developer fixing a critical bug, you don't care about saving $2. You care about the answer being right. "Free" is expensive if it gives you the wrong code.
For Leaders: Stop forcing your team to use cheap, inferior models. Give them the Ferrari. The productivity gains outweigh the subscription cost by orders of magnitude.
5. From Assistant to Collaborator (The Context Explosion)
In 20 months:
- Average Prompt Length: 1,500 --> 6,000+ tokens (4x increase).
- Average Output Length: 150 --> 400+ tokens (3x increase).
We aren't asking simple questions anymore. We are dumping entire project contexts, strategy documents, and code repositories into the context window. We are using AI as a Collaborator, not just a search engine.
The Jevons Paradox: As AI gets cheaper and smarter, we don't spend less. We use it more for harder problems, driving total costs up.
6. The Global Shift: Asia is Paying Up
Asia's share of AI spending jumped from 13% to 31% in 2025.
North America's share dropped below 50% for the first time.
Language Stats:
- English: 82.87%
- Simplified Chinese: 4.95%
- Russian: 2.47%
Chinese is the second most used language in global AI interactions. This signals that AI is no longer a "Western-only" game. Expect future models to optimize heavily for Chinese language performance.
Conclusion: The Dual Reality
The data paints a fascinating picture of our new reality.
On one screen, a developer is architecting a system with DeepSeek.
On the other screen, a writer is building a fantasy novel with LLaMA.
AI is simultaneously the ultimate Productivity Tool and the ultimate Escapist Fantasy.
The landscape is shifting fast. Today's leader might be tomorrow's runner-up. But for us—the users—this competition is pure gold. It means better tools, more choices, and more power at our fingertips.
Mercury Technology Solutions: Accelerate Digitality.
