Is China Winning the AI War? | State of AI 2025 by OpenRouter

2025-12-24 · Source: Venelin Valkov · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, medium

Summary

An Open Router report analyzes the real-world production usage of over 300 AI models, focusing on open-weight models and their task applications. The report, based on a 0.25% sample of prompts and responses categorized by Google T classifier between November 2024 and November 2025, reveals that agentic coding is the dominant use case, accounting for a significant increase in token usage from 10% to 50%. While closed models like those from Anthropic, Google, and OpenAI lead in overall token consumption, open-weight models, particularly from China (e.g., Deepseek R1, Kimik K2, Quentry), show steady adoption. The average prompt token count increased from 2,000 to 5,000, and completion tokens also rose, largely driven by agentic coding. Model stickiness is higher for early adopters, with Anthropic leading in agentic programming capabilities among major providers.

Key takeaway

For AI Architects evaluating model deployments, recognize that agentic coding is a primary driver of token usage and model stickiness. Your strategy should prioritize models excelling in programming tasks, such as Anthropic's offerings, and consider the long-term engagement benefits of early adoption. Be prepared for increased token consumption in prompts and completions when designing agentic systems.

Key insights

Agentic coding dominates AI model token usage, driving increased prompt and completion lengths in production environments.

Principles

• Early adoption correlates with higher model stickiness.
• Closed models lead in overall token consumption.
• Agentic systems drive reasoning token generation.

Method

Open Router categorizes a 0.25% sample of prompts/responses using Google T classifier to analyze model usage and task distribution, mapping classifications to categories like programming and roleplay.

In practice

• Prioritize agentic coding capabilities for AI model development.
• Focus on early user engagement for model adoption.
• Monitor token usage for cost and performance optimization.

Topics

• Agentic AI
• Open-Weight Models
• Token Usage Trends
• AI Model Adoption
• AI Application Domains

Best for: AI Architect, AI Product Manager, Entrepreneur, AI Engineer, Machine Learning Engineer, Prompt Engineer

Related on AIssential

• The Dawn of the Agent Age — The AI landscape rapidly shifted in January, marking the "agent era" with advanced agentic coding and a widening AI adoption gap.
• Agents Just Passed Humans in Token Usage. And They Burn Far More Than Anyone Budgeted. A Deep Dive With OpenRouter’s COO — Agentic AI usage now dominates, demanding robust inference, tool calling, and cost management.
• Dylan Patel of SemiAnalysis on the $200B AI CapEx, Chip Wars, and Why Google Might Have No Profits in 2027 — In-Context Cooking — Claude Code represents a significant shift in AI, enabling non-engineers to automate complex research and financial modeling tasks efficiently.
• Anthropic is about to become the first profitable AI lab — Anthropic is nearing profitability driven by coding tools and agentic AI, despite increased user costs and compute demands.
• OpenAI VS Anthropic — Frontier AI labs are rapidly advancing generative coding, focusing on agentic systems and self-improvement.
• Claude Code, Codex and Agentic Coding #7: Auto Mode — Auto Mode enhances agentic coding safety by selectively requiring human approval for high-risk actions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Venelin Valkov.