OpenRouter Analysis: 100 Trillion Token Research Report—What Are Humans Really Using AI For, The Rise of Chinese Models, and the Secret to User Retention

OpenRouter Report Reveals the Advent of the AI Agent Reasoning Era, the Rapid Rise of Chinese Open-Source Models Driven by Roleplay and Programming Capabilities in the Global Market, and Shares Real User Scenarios

(Background: ARK’s Cathie Wood Asserts “AI Is Not a Bubble”: Replicating the Internet’s Wealth Explosion Moment) (Additional Context: Google Officially Launches “Gemini 3”! The World’s Most Advanced AI Model—What’s New?)

This week, AI model aggregation platform OpenRouter released its highly anticipated annual report, “2025 AI State of the Union.” Unlike reports based on surveys or media hype, this analysis is grounded in empirical data from 100 trillion tokens generated on the OpenRouter platform.

Spanning over 300 large language models (LLMs) and 60+ providers, these data offer a fresh perspective on how AI is really being used in the real world.

Paradigm Shift: From “Next-Word Prediction” to “Agent Reasoning”

If the AI race before 2024 was about who could generate text more fluently, then the theme for 2025 is “thinking.” The report highlights the past year as a true watershed in LLM evolution.

o1 Model and the Beginning of the Reasoning Era The report defines December 5, 2024, as a key turning point when OpenAI released o1, the first widely adopted reasoning model. Prior to this, even the most advanced models (like Claude 3.5 Sonnet or GPT-4) were essentially probabilistic predictors performing single-pass, forward computations, mimicking reasoning based on training data patterns rather than conducting real internal computation.

The arrival of o1 changed everything. It introduced an internal computation process during inference—before delivering a final answer, the model engages in multi-step deliberation, latent planning, and self-correction.

Data Evidence: According to OpenRouter statistics, since early 2025, the share of tokens flowing to “reasoning-optimized models” has grown exponentially. By the end of 2025, over 50% of token usage involved models with reasoning capabilities. This signifies that users no longer see AI merely as chatbots, but as “brains” capable of solving complex logical problems.

The Rise of Agent Workflows With enhanced reasoning abilities, “agent workflows” have become another key term. Users are building complex automated systems where models don’t just answer questions but “take action.”

The report specifically analyzes “tool use” data—AI models’ ability to use external tools such as web search, code execution, or database queries.

Trend: Tool use saw a steady and significant rise throughout 2025.

Model Landscape: Initially, this space was dominated by OpenAI’s gpt-4o-mini and Anthropic’s Claude 3.5 series. But in the second half of the year, competition flourished: Google’s Gemini Flash, xAI’s Grok Code Fast, and China’s GLM 4.5 rapidly caught up in tool use capabilities, driving broader adoption of enterprise automation.

The Battle Between Open and Closed Source: The Rise of Chinese Models

On the ecosystem front, the report reveals a dual-track world: while closed-source models still dominate revenues, open-source models (OSS) play a key role in innovation and specialized domains.

Closed Source Still Dominates, but Open Source Is Surging While closed-source models (such as GPT-5, Claude 4.5) still account for about 70% of total token consumption, open-source models’ market share has grown sharply over the past year, now nearing 30%.

This indicates that developers and enterprises are increasingly seeking balance among controllability, privacy, and cost—high-quality open-source models provide this option.

The Dominance of Chinese Open-Source Models A striking finding in the report: Chinese-developed open-source models have become the backbone of the global open-source ecosystem.

Key Players: Alibaba’s Qwen series, DeepSeek, and Moonshot AI’s Kimi models account for a large proportion of OpenRouter’s global usage.

Competitive Advantages: The report notes China’s rapid rise is mainly due to “intensive iteration cycles” and “highly competitive model quality.” For example, Qwen 2.5 Coder’s performance in programming directly challenges many closed-source models’ dominance.

Unique Positioning: Data show that Chinese open-source models feature relatively less “content filtering” (compared to the strict safeguards of some Western models), making them especially popular for creative writing, roleplay, and story continuation scenarios.

The “M-shaped” Development of Model Sizes

There are also interesting structural changes within the open-source market:

Decline of Small Models: Despite Google’s launch of models like Gemma, overall usage of small models is declining, as users find their intelligence limits hamper application in complex tasks.

Golden Era for Medium-Sized Models: Models with around 32B parameters have become the “sweet spot.” Represented by Qwen 2.5 Coder 32B, these models strike a perfect balance between performance and inference cost, achieving true model-market fit.

Diversity in Large Models: In the 70B to 200B+ parameter range, the market is not dominated by a single model but features diverse competition (e.g., Llama 3 series, Qwen 72B). Users switch between large models based on specific tasks.

Real-World Scenarios: What Are Humans Actually Doing with AI?

Perhaps the most enlightening section of the report comes from content categorization of billions of requests using OpenRouter’s GoogleTagClassifier system, challenging the stereotype that “AI is mainly used for business productivity.”

Roleplay Data show roleplaying games are the largest use category for open-source models, accounting for over 60%.

Phenomenon: A large number of users are engaging in virtual character conversations, interactive fiction, and immersive gaming experiences. This is not just a niche hobby but a huge, essential market.

Drivers: Users prefer open-source models for roleplay mainly because these models (especially those from non-US labs) typically have fewer moderation restrictions, enabling richer and less constrained interactions.

Business Value: These users show remarkable stickiness, building deep emotional connections and usage habits around specific models.

Programming Closely following is programming—a category experiencing the fastest growth…