For the previous 12 months, we’ve been advised that synthetic intelligence is revolutionising productiveness—serving to us write emails, generate code, and summarise paperwork. However what if the actuality of how folks really use AI is utterly completely different from what we’ve been led to consider?
An information-driven study by OpenRouter has simply pulled again the curtain on real-world AI utilization by analysing over 100 trillion tokens—basically billions upon billions of conversations and interactions with giant language fashions like ChatGPT, Claude, and dozens of others. The findings problem many assumptions about the AI revolution.
OpenRouter is a multi-model AI inference platform that routes requests throughout greater than 300 fashions from over 60 suppliers—from OpenAI and Anthropic to open-source options like DeepSeek and Meta’s LLaMA.
With over 50% of its utilization originating exterior the United States and serving thousands and thousands of builders globally, the platform gives a novel cross-section of how AI is really deployed throughout completely different geographies, use circumstances, and person sorts.
Importantly, the examine analysed metadata from billions of interactions with out accessing the precise textual content of conversations, preserving person privateness whereas revealing behavioural patterns.
The roleplay revolution no one noticed coming
Maybe the most stunning discovery: greater than half of all open-source AI mannequin utilization isn’t for productiveness in any respect. It’s for roleplay and inventive storytelling.
Sure, you learn that proper. Whereas tech executives tout AI’s potential to rework enterprise, customers are spending the majority of their time partaking in character-driven conversations, interactive fiction, and gaming situations.
Over 50% of open-source mannequin interactions fall into this class, dwarfing even programming help.
“This counters an assumption that LLMs are largely used for writing code, emails, or summaries,” the report states. “In actuality, many customers interact with these fashions for companionship or exploration.”
This isn’t simply informal chatting. The information exhibits customers deal with AI fashions as structured roleplaying engines, with 60% of roleplay tokens falling underneath particular gaming situations and inventive writing contexts. It’s a large, largely invisible use case that’s reshaping how AI firms take into consideration their merchandise.
Programming’s meteoric rise
Whereas roleplay dominates open-source utilization, programming has grow to be the fastest-growing class throughout all AI fashions. At the begin of 2025, coding-related queries accounted for simply 11% of complete AI utilization. By the finish of the 12 months, that determine had exploded to over 50%.
This development displays AI’s deepening integration into software program growth. Common immediate lengths for programming duties have grown fourfold, from round 1,500 tokens to over 6,000, with some code-related requests exceeding 20,000 tokens—roughly equal to feeding a complete codebase into an AI mannequin for evaluation.
For context, programming queries now generate a few of the longest and most advanced interactions in the whole AI ecosystem. Builders aren’t simply asking for easy code snippets anymore; they’re conducting subtle debugging classes, architectural evaluations, and multi-step downside fixing.
Anthropic’s Claude fashions dominate this house, capturing over 60% of programming-related utilization for many of 2025, although competitors is intensifying as Google, OpenAI, and open-source options achieve floor.
The Chinese language AI surge
One other main revelation: Chinese language AI fashions now account for roughly 30% of worldwide utilization—practically triple their 13% share at the begin of 2025.
Fashions from DeepSeek, Qwen (Alibaba), and Moonshot AI have quickly gained traction, with DeepSeek alone processing 14.37 trillion tokens throughout the examine interval. This represents a elementary shift in the world AI panorama, the place Western firms not maintain unchallenged dominance.
Simplified Chinese language is now the second-most frequent language for AI interactions globally at 5% of complete utilization, behind solely English at 83%. Asia’s total share of AI spending greater than doubled from 13% to 31%, with Singapore rising as the second-largest nation by utilization after the United States.
The rise of “Agentic” AI
The examine introduces an idea that can outline AI’s subsequent part: agentic inference. This means AI fashions are not simply answering single questions—they’re executing multi-step duties, calling external instruments, and reasoning throughout prolonged conversations.
The share of AI interactions labeled as “reasoning-optimised” jumped from practically zero in early 2025 to over 50% by 12 months’s finish. This displays a elementary shift from AI as a textual content generator to AI as an autonomous agent able to planning and execution.
“The median LLM request is not a easy query or remoted instruction,” the researchers clarify. “As an alternative, it is a part of a structured, agent-like loop, invoking external instruments, reasoning over state, and persisting throughout longer contexts.”
Consider it this manner: as a substitute of asking AI to “write a operate,” you’re now asking it to “debug this codebase, determine the efficiency bottleneck, and implement an answer”—and it could possibly really do it.
The “Glass Slipper Impact”
One in every of the examine’s most fascinating insights relates to person retention. Researchers found what they name the Cinderella “Glass Slipper” impact—a phenomenon the place AI fashions that are “first to remedy” a crucial downside create lasting person loyalty.
When a newly launched mannequin completely matches a beforehand unmet want—the metaphorical “glass slipper”—these early customers stick round far longer than later adopters. For instance, the June 2025 cohort of Google’s Gemini 2.5 Professional retained roughly 40% of customers at month 5, considerably greater than later cohorts.
This challenges typical knowledge about AI competitors. Being first issues, however particularly being first to remedy a high-value downside creates a sturdy aggressive benefit. Customers embed these fashions into their workflows, making switching pricey each technically and behaviorally.
Value doesn’t matter (as a lot as you’d suppose)
Maybe counterintuitively, the examine reveals that AI utilization is comparatively price-inelastic. A ten% lower in value corresponds to solely a few 0.5-0.7% enhance in utilization.
Premium fashions from Anthropic and OpenAI command $2-35 per million tokens whereas sustaining excessive utilization, whereas finances choices like DeepSeek and Google’s Gemini Flash obtain comparable scale at underneath $0.40 per million tokens. Each coexist efficiently.
“The LLM market does not appear to behave like a commodity simply but,” the report concludes. “Customers steadiness price with reasoning high quality, reliability, and breadth of functionality.”
This means AI hasn’t grow to be a race to the backside on pricing. High quality, reliability, and functionality nonetheless command premiums—no less than for now.
What this implies going ahead
The OpenRouter examine paints an image of real-world AI utilization that’s much more nuanced than business narratives counsel. Sure, AI is reworking programming {and professional} work. However it’s additionally creating completely new classes of human-computer interplay by roleplay and inventive functions.
The market is diversifying geographically, with China rising as a significant power. The know-how is evolving from easy textual content technology to advanced, multi-step reasoning. And person loyalty relies upon much less on being first to market than on being first to actually remedy an issue.
As the report notes, “methods during which folks use LLMs do not all the time align with expectations and range considerably nation by nation, state by state, use case by use case.”
Understanding these real-world patterns—not simply benchmark scores or advertising claims—shall be essential as AI turns into additional embedded in every day life. The hole between how we expect AI is used and the way it’s really used is wider than most realise. This examine helps shut that hole.
See additionally: Deep Cogito v2: Open-source AI that hones its reasoning skills
Need to be taught extra about AI and large information from business leaders? Take a look at AI & Big Data Expo going down in Amsterdam, California, and London. The great occasion is a part of TechEx and is co-located with different main know-how occasions together with the Cyber Security Expo. Click on here for extra information.
AI Information is powered by TechForge Media. Discover different upcoming enterprise know-how occasions and webinars here.
Disclaimer: This article is sourced from external platforms. OverBeta has not independently verified the information. Readers are advised to verify details before relying on them.
