Chinese-built artificial intelligence models now account for 61% of all AI usage on OpenRouter, the world's largest AI model marketplace, a dramatic shift from less than 1% just 18 months ago. This rapid market takeover reflects a fundamental change in how developers choose AI tools: performance has become nearly equal between Chinese and American models, but pricing has not, making cost the decisive factor.

How Did Chinese AI Models Capture the Market So Quickly?

The story begins in January 2025, when a Chinese AI lab called DeepSeek released a model that matched the performance of OpenAI's best offerings at a fraction of the cost. While American frontier models cost thousands of dollars to run through standard evaluation tests, DeepSeek completed the same work for a few hundred dollars. Developers took notice immediately.

Throughout 2025, a wave of Chinese AI laboratories pushed out increasingly competitive models. Moonshot AI, maker of the Kimi K2.5 model, joined competitors including MiniMax, Zhipu AI, Alibaba's Qwen team, and Xiaomi in releasing models that combined serious performance with aggressive pricing. These were not cheap knock-offs; benchmarks showed them trading blows with GPT-4 and Claude on coding tasks, reasoning challenges, and the kind of multi-step "agentic" work that companies actually use AI for in production.

The cost gap tells the real story. Running Anthropic's Claude through ten standard evaluation tasks costs around $4,811. Running DeepSeek through the same tasks costs $1,071. Zhipu's GLM model costs $544 for the same work. That represents roughly a nine-to-one price gap for comparable output. For individual developers, that is a curiosity. For enterprise companies running millions of AI calls per day, that is a budget-breaking difference.

What Do the Latest Usage Numbers Actually Show?

OpenRouter's own usage data, publicly tracked and updated weekly, reveals the scale of the shift. As of February 2026, models built in China account for 61% of total token consumption on the platform. The top three most-used models were all developed by Chinese AI laboratories, with aggregate consumption reaching 5.3 trillion tokens out of 8.7 trillion total tokens consumed by the top ten models.

MiniMax M2.5 topped the usage charts with 2.45 trillion tokens, followed by Kimi K2.5 and Zhipu GLM-5, all Chinese models. On software engineering tasks, MiniMax M2.5 scored 80.2%, nearly identical to a leading American frontier model at 80.8%, yet the price gap was enormous: $0.30 per million tokens versus $5.00, roughly 17 times cheaper.

Programming has evolved into the largest category of token usage on OpenRouter, expanding from 11% to over 50% of the total throughout 2025. Agent-driven workflows, which are exactly the use cases where Chinese models have optimized hardest, now generate more than half of all output tokens on the platform.

Why Are American Companies Switching to Chinese AI Models?

The shift reflects rational economic decision-making by developers and companies. OpenRouter's Chief Operating Officer stated that Chinese open-weight models have captured significant market share because they are "disproportionately heavy in agentic flows run by U.S. firms." American companies in San Francisco, New York, and Chicago are running their AI workflows on Chinese models, not because they are forced to, but because it makes financial sense.

Operating Officer

The impact showed up in earnings calls this season. Meta, Shopify, Spotify, and Pinterest all flagged rising AI and inference costs as a drag on margins. Companies are looking for cheaper options, and Chinese AI is that cheaper option. A CloudZero survey found the share of companies spending at least $100,000 a month on AI rose to 45% in 2025, from 20% in the same period a year earlier. As AI budgets exploded, the pressure to find cost-effective solutions became intense.

Steps to Understand the Market Shift in AI Pricing and Performance

• Compare Performance Benchmarks: Look at how different AI models score on standard evaluation tasks like coding, reasoning, and multi-step problem-solving. Chinese models like Kimi K2.5 and MiniMax M2.5 now achieve nearly identical scores to American models like Claude and GPT-4.
• Calculate Total Cost of Ownership: Beyond the per-token price, consider how much you will spend monthly on AI inference across your entire organization. A 17-fold price difference compounds quickly when running millions of API calls per day.
• Evaluate Use Case Fit: Determine whether your primary AI workload is programming, agent-driven automation, or general-purpose tasks. Chinese models have optimized specifically for agentic workflows, where the price advantage is most pronounced.
• Monitor Market Share Trends: Track which models are gaining adoption on platforms like OpenRouter. Usage data reflects real developer choices and can signal which models are becoming industry standards.

What Does This Mean for OpenAI and Anthropic's IPO Plans?

The timing could not be worse for OpenAI and Anthropic, both of which are preparing some of the most anticipated stock market listings in tech history. OpenAI is reportedly preparing for a potential IPO in late 2026 with a target valuation of roughly $1 trillion. Anthropic was valued at around $380 billion in February 2026 and is seeking its next round at between $600 billion and $900 billion.

Both valuations rest on a specific assumption: that these companies control the market. The data from OpenRouter suggests they do not anymore. The performance gap between Chinese and American models has essentially closed. The price gap has not. And developers, who are rational and budget-conscious, have responded accordingly by shifting their usage to cheaper alternatives.

This market shift represents not a slow drift but what the data describes as a structural collapse of American dominance in one of the most important technology markets of the century. From 1% to 60% in roughly eighteen months is not a market trend; it is a market takeover.